unet_2d_blocks.py 127 KB
Newer Older
Patrick von Platen's avatar
Patrick von Platen committed
1
# Copyright 2023 The HuggingFace Team. All rights reserved.
Patrick von Platen's avatar
Patrick von Platen committed
2
3
4
5
6
7
8
9
10
11
12
13
#
# 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.
14
from typing import Any, Dict, Optional, Tuple
15

16
import numpy as np
17
import torch
18
import torch.nn.functional as F
Patrick von Platen's avatar
Patrick von Platen committed
19
20
from torch import nn

21
from ..utils import is_torch_version, logging
22
from ..utils.torch_utils import apply_freeu
23
from .activations import get_activation
24
from .attention import AdaGroupNorm
25
from .attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
26
from .dual_transformer_2d import DualTransformer2DModel
27
from .resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D
28
from .transformer_2d import Transformer2DModel
Patrick von Platen's avatar
Patrick von Platen committed
29
30


31
32
33
logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


34
35
36
37
38
39
40
41
42
def get_down_block(
    down_block_type,
    num_layers,
    in_channels,
    out_channels,
    temb_channels,
    add_downsample,
    resnet_eps,
    resnet_act_fn,
43
    transformer_layers_per_block=1,
44
    num_attention_heads=None,
45
    resnet_groups=None,
46
    cross_attention_dim=None,
Patrick von Platen's avatar
Patrick von Platen committed
47
    downsample_padding=None,
48
    dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
49
    use_linear_projection=False,
50
    only_cross_attention=False,
51
    upcast_attention=False,
Will Berman's avatar
Will Berman committed
52
    resnet_time_scale_shift="default",
53
    attention_type="default",
54
55
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
56
    cross_attention_norm=None,
57
    attention_head_dim=None,
58
    downsample_type=None,
59
    dropout=0.0,
60
):
61
62
63
64
65
66
67
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warn(
            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
        )
        attention_head_dim = num_attention_heads

Patrick von Platen's avatar
Patrick von Platen committed
68
69
70
    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
    if down_block_type == "DownBlock2D":
        return DownBlock2D(
71
72
73
74
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
75
            dropout=dropout,
76
77
78
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
79
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
80
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
81
82
83
84
85
86
87
88
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "ResnetDownsampleBlock2D":
        return ResnetDownsampleBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
89
            dropout=dropout,
Will Berman's avatar
Will Berman committed
90
91
92
93
94
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
95
96
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
97
        )
Patrick von Platen's avatar
Patrick von Platen committed
98
    elif down_block_type == "AttnDownBlock2D":
99
100
101
102
        if add_downsample is False:
            downsample_type = None
        else:
            downsample_type = downsample_type or "conv"  # default to 'conv'
Patrick von Platen's avatar
Patrick von Platen committed
103
        return AttnDownBlock2D(
104
105
106
107
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
108
            dropout=dropout,
109
110
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
111
            resnet_groups=resnet_groups,
112
            downsample_padding=downsample_padding,
113
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
114
            resnet_time_scale_shift=resnet_time_scale_shift,
115
            downsample_type=downsample_type,
116
        )
Patrick von Platen's avatar
Patrick von Platen committed
117
    elif down_block_type == "CrossAttnDownBlock2D":
118
        if cross_attention_dim is None:
119
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
Patrick von Platen's avatar
Patrick von Platen committed
120
        return CrossAttnDownBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
121
            num_layers=num_layers,
122
            transformer_layers_per_block=transformer_layers_per_block,
Patrick von Platen's avatar
Patrick von Platen committed
123
124
125
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
126
            dropout=dropout,
Patrick von Platen's avatar
Patrick von Platen committed
127
128
129
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
130
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
131
            downsample_padding=downsample_padding,
132
            cross_attention_dim=cross_attention_dim,
133
            num_attention_heads=num_attention_heads,
134
            dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
135
            use_linear_projection=use_linear_projection,
136
            only_cross_attention=only_cross_attention,
137
            upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
138
            resnet_time_scale_shift=resnet_time_scale_shift,
139
            attention_type=attention_type,
Will Berman's avatar
Will Berman committed
140
141
142
143
144
145
146
147
148
        )
    elif down_block_type == "SimpleCrossAttnDownBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
        return SimpleCrossAttnDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
149
            dropout=dropout,
Will Berman's avatar
Will Berman committed
150
151
152
153
154
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
155
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
156
            resnet_time_scale_shift=resnet_time_scale_shift,
157
158
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
159
            only_cross_attention=only_cross_attention,
160
            cross_attention_norm=cross_attention_norm,
Patrick von Platen's avatar
Patrick von Platen committed
161
        )
Patrick von Platen's avatar
Patrick von Platen committed
162
163
    elif down_block_type == "SkipDownBlock2D":
        return SkipDownBlock2D(
164
165
166
167
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
168
            dropout=dropout,
169
170
171
172
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
173
            resnet_time_scale_shift=resnet_time_scale_shift,
174
        )
Patrick von Platen's avatar
Patrick von Platen committed
175
176
    elif down_block_type == "AttnSkipDownBlock2D":
        return AttnSkipDownBlock2D(
177
178
179
180
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
181
            dropout=dropout,
182
183
184
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
185
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
186
            resnet_time_scale_shift=resnet_time_scale_shift,
187
        )
188
189
190
191
192
    elif down_block_type == "DownEncoderBlock2D":
        return DownEncoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
193
            dropout=dropout,
194
195
196
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
197
            resnet_groups=resnet_groups,
198
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
199
            resnet_time_scale_shift=resnet_time_scale_shift,
200
        )
Will Berman's avatar
Will Berman committed
201
202
203
204
205
    elif down_block_type == "AttnDownEncoderBlock2D":
        return AttnDownEncoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
206
            dropout=dropout,
Will Berman's avatar
Will Berman committed
207
208
209
210
211
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
212
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
213
            resnet_time_scale_shift=resnet_time_scale_shift,
Will Berman's avatar
Will Berman committed
214
        )
215
216
217
218
219
220
    elif down_block_type == "KDownBlock2D":
        return KDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
221
            dropout=dropout,
222
223
224
225
226
227
228
229
230
231
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
        )
    elif down_block_type == "KCrossAttnDownBlock2D":
        return KCrossAttnDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
232
            dropout=dropout,
233
234
235
236
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            cross_attention_dim=cross_attention_dim,
237
            attention_head_dim=attention_head_dim,
238
239
            add_self_attention=True if not add_downsample else False,
        )
Will Berman's avatar
Will Berman committed
240
    raise ValueError(f"{down_block_type} does not exist.")
241
242
243
244
245
246


def get_up_block(
    up_block_type,
    num_layers,
    in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
247
248
    out_channels,
    prev_output_channel,
249
250
251
252
    temb_channels,
    add_upsample,
    resnet_eps,
    resnet_act_fn,
253
    resolution_idx=None,
254
    transformer_layers_per_block=1,
255
    num_attention_heads=None,
256
    resnet_groups=None,
257
    cross_attention_dim=None,
258
    dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
259
    use_linear_projection=False,
260
    only_cross_attention=False,
261
    upcast_attention=False,
Will Berman's avatar
Will Berman committed
262
    resnet_time_scale_shift="default",
263
    attention_type="default",
264
265
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
266
    cross_attention_norm=None,
267
    attention_head_dim=None,
268
    upsample_type=None,
269
    dropout=0.0,
270
):
271
272
273
274
275
276
277
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warn(
            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
        )
        attention_head_dim = num_attention_heads

Patrick von Platen's avatar
Patrick von Platen committed
278
279
280
    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
    if up_block_type == "UpBlock2D":
        return UpBlock2D(
281
282
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
283
284
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
285
            temb_channels=temb_channels,
286
            resolution_idx=resolution_idx,
287
            dropout=dropout,
288
289
290
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
291
            resnet_groups=resnet_groups,
Will Berman's avatar
Will Berman committed
292
293
294
295
296
297
298
299
300
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "ResnetUpsampleBlock2D":
        return ResnetUpsampleBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
301
            resolution_idx=resolution_idx,
302
            dropout=dropout,
Will Berman's avatar
Will Berman committed
303
304
305
306
307
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
308
309
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
310
        )
Patrick von Platen's avatar
Patrick von Platen committed
311
    elif up_block_type == "CrossAttnUpBlock2D":
312
313
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
Patrick von Platen's avatar
Patrick von Platen committed
314
        return CrossAttnUpBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
315
            num_layers=num_layers,
316
            transformer_layers_per_block=transformer_layers_per_block,
Patrick von Platen's avatar
Patrick von Platen committed
317
318
319
320
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
321
            resolution_idx=resolution_idx,
322
            dropout=dropout,
Patrick von Platen's avatar
Patrick von Platen committed
323
324
325
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
326
            resnet_groups=resnet_groups,
327
            cross_attention_dim=cross_attention_dim,
328
            num_attention_heads=num_attention_heads,
329
            dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
330
            use_linear_projection=use_linear_projection,
331
            only_cross_attention=only_cross_attention,
332
            upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
333
            resnet_time_scale_shift=resnet_time_scale_shift,
334
            attention_type=attention_type,
Will Berman's avatar
Will Berman committed
335
336
337
338
339
340
341
342
343
344
        )
    elif up_block_type == "SimpleCrossAttnUpBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
        return SimpleCrossAttnUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
345
            resolution_idx=resolution_idx,
346
            dropout=dropout,
Will Berman's avatar
Will Berman committed
347
348
349
350
351
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
352
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
353
            resnet_time_scale_shift=resnet_time_scale_shift,
354
355
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
356
            only_cross_attention=only_cross_attention,
357
            cross_attention_norm=cross_attention_norm,
Patrick von Platen's avatar
Patrick von Platen committed
358
        )
Patrick von Platen's avatar
Patrick von Platen committed
359
    elif up_block_type == "AttnUpBlock2D":
360
361
362
363
364
        if add_upsample is False:
            upsample_type = None
        else:
            upsample_type = upsample_type or "conv"  # default to 'conv'

Patrick von Platen's avatar
Patrick von Platen committed
365
        return AttnUpBlock2D(
366
367
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
368
369
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
370
            temb_channels=temb_channels,
371
            resolution_idx=resolution_idx,
372
            dropout=dropout,
373
374
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
375
            resnet_groups=resnet_groups,
376
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
377
            resnet_time_scale_shift=resnet_time_scale_shift,
378
            upsample_type=upsample_type,
379
        )
Patrick von Platen's avatar
Patrick von Platen committed
380
381
    elif up_block_type == "SkipUpBlock2D":
        return SkipUpBlock2D(
382
383
384
385
386
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
387
            resolution_idx=resolution_idx,
388
            dropout=dropout,
389
390
391
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
Will Berman's avatar
Will Berman committed
392
            resnet_time_scale_shift=resnet_time_scale_shift,
393
        )
Patrick von Platen's avatar
Patrick von Platen committed
394
395
    elif up_block_type == "AttnSkipUpBlock2D":
        return AttnSkipUpBlock2D(
396
397
398
399
400
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
401
            resolution_idx=resolution_idx,
402
            dropout=dropout,
403
404
405
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
406
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
407
            resnet_time_scale_shift=resnet_time_scale_shift,
408
        )
409
410
411
412
413
    elif up_block_type == "UpDecoderBlock2D":
        return UpDecoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
414
            resolution_idx=resolution_idx,
415
            dropout=dropout,
416
417
418
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
419
            resnet_groups=resnet_groups,
Will Berman's avatar
Will Berman committed
420
            resnet_time_scale_shift=resnet_time_scale_shift,
YiYi Xu's avatar
YiYi Xu committed
421
            temb_channels=temb_channels,
422
        )
Will Berman's avatar
Will Berman committed
423
424
425
426
427
    elif up_block_type == "AttnUpDecoderBlock2D":
        return AttnUpDecoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
428
            resolution_idx=resolution_idx,
429
            dropout=dropout,
Will Berman's avatar
Will Berman committed
430
431
432
433
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
434
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
435
            resnet_time_scale_shift=resnet_time_scale_shift,
YiYi Xu's avatar
YiYi Xu committed
436
            temb_channels=temb_channels,
Will Berman's avatar
Will Berman committed
437
        )
438
439
440
441
442
443
    elif up_block_type == "KUpBlock2D":
        return KUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
444
            resolution_idx=resolution_idx,
445
            dropout=dropout,
446
447
448
449
450
451
452
453
454
455
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
        )
    elif up_block_type == "KCrossAttnUpBlock2D":
        return KCrossAttnUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
456
            resolution_idx=resolution_idx,
457
            dropout=dropout,
458
459
460
461
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            cross_attention_dim=cross_attention_dim,
462
            attention_head_dim=attention_head_dim,
463
464
        )

465
    raise ValueError(f"{up_block_type} does not exist.")
466
467


468
class AutoencoderTinyBlock(nn.Module):
469
    """
Patrick von Platen's avatar
Patrick von Platen committed
470
471
    Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
    blocks.
472
473
474
475

    Args:
        in_channels (`int`): The number of input channels.
        out_channels (`int`): The number of output channels.
Patrick von Platen's avatar
Patrick von Platen committed
476
477
        act_fn (`str`):
            ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
478
479

    Returns:
Patrick von Platen's avatar
Patrick von Platen committed
480
481
        `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
        `out_channels`.
482
483
    """

484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
    def __init__(self, in_channels: int, out_channels: int, act_fn: str):
        super().__init__()
        act_fn = get_activation(act_fn)
        self.conv = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
            act_fn,
            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
            act_fn,
            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
        )
        self.skip = (
            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
            if in_channels != out_channels
            else nn.Identity()
        )
        self.fuse = nn.ReLU()

    def forward(self, x):
        return self.fuse(self.conv(x) + self.skip(x))


Patrick von Platen's avatar
Patrick von Platen committed
505
class UNetMidBlock2D(nn.Module):
506
507
508
509
510
511
512
513
514
    """
    A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.

    Args:
        in_channels (`int`): The number of input channels.
        temb_channels (`int`): The number of temporal embedding channels.
        dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
        num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
        resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
Patrick von Platen's avatar
Patrick von Platen committed
515
516
517
        resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
            The type of normalization to apply to the time embeddings. This can help to improve the performance of the
            model on tasks with long-range temporal dependencies.
518
        resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
Patrick von Platen's avatar
Patrick von Platen committed
519
520
        resnet_groups (`int`, *optional*, defaults to 32):
            The number of groups to use in the group normalization layers of the resnet blocks.
521
        attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
Patrick von Platen's avatar
Patrick von Platen committed
522
523
        resnet_pre_norm (`bool`, *optional*, defaults to `True`):
            Whether to use pre-normalization for the resnet blocks.
524
        add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
Patrick von Platen's avatar
Patrick von Platen committed
525
526
527
        attention_head_dim (`int`, *optional*, defaults to 1):
            Dimension of a single attention head. The number of attention heads is determined based on this value and
            the number of input channels.
528
529
530
        output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.

    Returns:
Patrick von Platen's avatar
Patrick von Platen committed
531
532
        `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
        in_channels, height, width)`.
533
534
535

    """

Patrick von Platen's avatar
Patrick von Platen committed
536
537
538
539
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
540
        dropout: float = 0.0,
541
        num_layers: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
542
        resnet_eps: float = 1e-6,
YiYi Xu's avatar
YiYi Xu committed
543
        resnet_time_scale_shift: str = "default",  # default, spatial
Patrick von Platen's avatar
Patrick von Platen committed
544
545
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
546
        attn_groups: Optional[int] = None,
547
        resnet_pre_norm: bool = True,
Will Berman's avatar
Will Berman committed
548
        add_attention: bool = True,
549
        attention_head_dim=1,
Patrick von Platen's avatar
Patrick von Platen committed
550
551
552
        output_scale_factor=1.0,
    ):
        super().__init__()
553
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
Will Berman's avatar
Will Berman committed
554
        self.add_attention = add_attention
Patrick von Platen's avatar
Patrick von Platen committed
555

556
557
558
        if attn_groups is None:
            attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None

559
560
        # there is always at least one resnet
        resnets = [
561
            ResnetBlock2D(
562
563
564
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
565
                eps=resnet_eps,
566
567
568
569
570
571
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
Patrick von Platen's avatar
Patrick von Platen committed
572
            )
573
574
        ]
        attentions = []
Patrick von Platen's avatar
Patrick von Platen committed
575

576
577
578
579
580
581
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
            )
            attention_head_dim = in_channels

582
        for _ in range(num_layers):
Will Berman's avatar
Will Berman committed
583
584
            if self.add_attention:
                attentions.append(
585
                    Attention(
Will Berman's avatar
Will Berman committed
586
                        in_channels,
587
588
                        heads=in_channels // attention_head_dim,
                        dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
589
590
                        rescale_output_factor=output_scale_factor,
                        eps=resnet_eps,
591
                        norm_num_groups=attn_groups,
YiYi Xu's avatar
YiYi Xu committed
592
                        spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
593
594
595
596
                        residual_connection=True,
                        bias=True,
                        upcast_softmax=True,
                        _from_deprecated_attn_block=True,
Will Berman's avatar
Will Berman committed
597
                    )
598
                )
Will Berman's avatar
Will Berman committed
599
600
601
            else:
                attentions.append(None)

602
            resnets.append(
603
                ResnetBlock2D(
604
605
606
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
607
                    eps=resnet_eps,
608
609
610
611
612
613
614
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
Patrick von Platen's avatar
Patrick von Platen committed
615
616
            )

617
618
619
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

Will Berman's avatar
Will Berman committed
620
    def forward(self, hidden_states, temb=None):
Patrick von Platen's avatar
Patrick von Platen committed
621
        hidden_states = self.resnets[0](hidden_states, temb)
622
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
Will Berman's avatar
Will Berman committed
623
            if attn is not None:
YiYi Xu's avatar
YiYi Xu committed
624
                hidden_states = attn(hidden_states, temb=temb)
Patrick von Platen's avatar
Patrick von Platen committed
625
            hidden_states = resnet(hidden_states, temb)
Patrick von Platen's avatar
Patrick von Platen committed
626

627
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
628

629

Patrick von Platen's avatar
Patrick von Platen committed
630
631
632
633
634
635
636
class UNetMidBlock2DCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
637
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
638
639
640
641
642
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
643
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
644
645
        output_scale_factor=1.0,
        cross_attention_dim=1280,
646
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
647
        use_linear_projection=False,
648
        upcast_attention=False,
649
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
650
651
652
    ):
        super().__init__()

653
        self.has_cross_attention = True
654
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
655
656
657
658
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

        # there is always at least one resnet
        resnets = [
659
            ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
            )
        ]
        attentions = []

        for _ in range(num_layers):
675
676
677
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
678
679
                        num_attention_heads,
                        in_channels // num_attention_heads,
680
                        in_channels=in_channels,
681
                        num_layers=transformer_layers_per_block,
682
683
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
684
                        use_linear_projection=use_linear_projection,
685
                        upcast_attention=upcast_attention,
686
                        attention_type=attention_type,
687
688
689
690
691
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
692
693
                        num_attention_heads,
                        in_channels // num_attention_heads,
694
695
696
697
698
                        in_channels=in_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
Patrick von Platen's avatar
Patrick von Platen committed
699
700
                )
            resnets.append(
701
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

718
719
        self.gradient_checkpointing = False

720
    def forward(
721
722
723
724
725
726
727
728
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ) -> torch.FloatTensor:
729
730
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
Patrick von Platen's avatar
Patrick von Platen committed
731
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
732
733
734
735
736
737
738
739
740
741
742
743
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
ethansmith2000's avatar
ethansmith2000 committed
744
                hidden_states = attn(
745
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
746
747
748
749
750
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
                )[0]
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
            else:
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
                )[0]
767
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
Will Berman's avatar
Will Berman committed
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783

        return hidden_states


class UNetMidBlock2DSimpleCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
784
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
785
786
        output_scale_factor=1.0,
        cross_attention_dim=1280,
787
        skip_time_act=False,
788
        only_cross_attention=False,
789
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
790
791
792
793
794
    ):
        super().__init__()

        self.has_cross_attention = True

795
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
796
797
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

798
        self.num_heads = in_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
799
800
801
802
803
804
805
806
807
808
809
810
811
812

        # there is always at least one resnet
        resnets = [
            ResnetBlock2D(
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
813
                skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
814
815
816
817
818
            )
        ]
        attentions = []

        for _ in range(num_layers):
819
820
821
822
            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
823
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
824
                Attention(
Will Berman's avatar
Will Berman committed
825
826
827
                    query_dim=in_channels,
                    cross_attention_dim=in_channels,
                    heads=self.num_heads,
828
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
829
830
831
832
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
833
                    only_cross_attention=only_cross_attention,
834
                    cross_attention_norm=cross_attention_norm,
835
                    processor=processor,
Will Berman's avatar
Will Berman committed
836
837
838
839
840
841
842
843
844
845
846
847
848
849
                )
            )
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
850
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
851
852
853
854
855
856
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

857
    def forward(
858
859
860
861
862
863
864
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
865
866
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
867
        lora_scale = cross_attention_kwargs.get("scale", 1.0)
868
869
870
871
872
873
874
875
876
877
878
879

        if attention_mask is None:
            # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
            mask = None if encoder_hidden_states is None else encoder_attention_mask
        else:
            # when attention_mask is defined: we don't even check for encoder_attention_mask.
            # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
            # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
            #       then we can simplify this whole if/else block to:
            #         mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
            mask = attention_mask

880
        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
Will Berman's avatar
Will Berman committed
881
882
883
884
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            # attn
            hidden_states = attn(
                hidden_states,
885
                encoder_hidden_states=encoder_hidden_states,
886
                attention_mask=mask,
887
                **cross_attention_kwargs,
Will Berman's avatar
Will Berman committed
888
889
890
            )

            # resnet
891
            hidden_states = resnet(hidden_states, temb, scale=lora_scale)
Patrick von Platen's avatar
Patrick von Platen committed
892
893
894
895

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
896
class AttnDownBlock2D(nn.Module):
897
898
899
900
901
902
903
904
905
906
907
908
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
909
        attention_head_dim=1,
910
        output_scale_factor=1.0,
911
        downsample_padding=1,
912
        downsample_type="conv",
913
914
915
916
    ):
        super().__init__()
        resnets = []
        attentions = []
917
        self.downsample_type = downsample_type
918

919
920
921
922
923
924
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

925
926
927
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
928
                ResnetBlock2D(
929
930
931
932
933
934
935
936
937
938
939
940
941
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
942
                Attention(
943
                    out_channels,
944
945
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
946
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
947
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
948
                    norm_num_groups=resnet_groups,
949
950
951
952
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
953
954
955
956
957
958
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

959
        if downsample_type == "conv":
960
            self.downsamplers = nn.ModuleList(
961
962
                [
                    Downsample2D(
963
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
964
965
                    )
                ]
966
            )
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
        elif downsample_type == "resnet":
            self.downsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
                        down=True,
                    )
                ]
            )
985
986
987
        else:
            self.downsamplers = None

988
989
990
991
992
    def forward(self, hidden_states, temb=None, upsample_size=None, cross_attention_kwargs=None):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}

        lora_scale = cross_attention_kwargs.get("scale", 1.0)

993
994
995
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
996
997
998
            cross_attention_kwargs.update({"scale": lora_scale})
            hidden_states = resnet(hidden_states, temb, scale=lora_scale)
            hidden_states = attn(hidden_states, **cross_attention_kwargs)
999
            output_states = output_states + (hidden_states,)
1000
1001
1002

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1003
                if self.downsample_type == "resnet":
1004
                    hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale)
1005
                else:
1006
                    hidden_states = downsampler(hidden_states, scale=lora_scale)
1007
1008
1009
1010
1011
1012

            output_states += (hidden_states,)

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
1013
class CrossAttnDownBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
1014
1015
1016
1017
1018
1019
1020
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
1021
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
1022
1023
1024
1025
1026
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
1027
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
1028
1029
1030
1031
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
1032
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
1033
        use_linear_projection=False,
1034
        only_cross_attention=False,
1035
        upcast_attention=False,
1036
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
1037
1038
1039
1040
1041
    ):
        super().__init__()
        resnets = []
        attentions = []

1042
        self.has_cross_attention = True
1043
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
1044
1045
1046
1047

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1048
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
1061
1062
1063
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
1064
1065
                        num_attention_heads,
                        out_channels // num_attention_heads,
1066
                        in_channels=out_channels,
1067
                        num_layers=transformer_layers_per_block,
1068
1069
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
1070
                        use_linear_projection=use_linear_projection,
1071
                        only_cross_attention=only_cross_attention,
1072
                        upcast_attention=upcast_attention,
1073
                        attention_type=attention_type,
1074
1075
1076
1077
1078
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
1079
1080
                        num_attention_heads,
                        out_channels // num_attention_heads,
1081
1082
1083
1084
1085
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
Patrick von Platen's avatar
Patrick von Platen committed
1086
1087
1088
1089
1090
1091
1092
1093
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1094
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
1095
1096
1097
1098
1099
1100
                    )
                ]
            )
        else:
            self.downsamplers = None

1101
1102
        self.gradient_checkpointing = False

1103
    def forward(
1104
1105
1106
1107
1108
1109
1110
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
Will Berman's avatar
Will Berman committed
1111
        additional_residuals=None,
1112
    ):
Patrick von Platen's avatar
Patrick von Platen committed
1113
1114
        output_states = ()

1115
1116
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0

Will Berman's avatar
Will Berman committed
1117
1118
1119
        blocks = list(zip(self.resnets, self.attentions))

        for i, (resnet, attn) in enumerate(blocks):
1120
1121
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
1122
                def create_custom_forward(module, return_dict=None):
1123
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
1124
1125
1126
1127
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
1128
1129
1130

                    return custom_forward

1131
1132
1133
1134
1135
1136
1137
                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
ethansmith2000's avatar
ethansmith2000 committed
1138
                hidden_states = attn(
1139
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
1140
1141
1142
1143
1144
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
1145
                )[0]
1146
            else:
1147
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
1148
1149
1150
1151
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
1152
1153
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
1154
1155
                    return_dict=False,
                )[0]
1156

Will Berman's avatar
Will Berman committed
1157
1158
1159
1160
            # apply additional residuals to the output of the last pair of resnet and attention blocks
            if i == len(blocks) - 1 and additional_residuals is not None:
                hidden_states = hidden_states + additional_residuals

1161
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
1162
1163
1164

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1165
                hidden_states = downsampler(hidden_states, scale=lora_scale)
Patrick von Platen's avatar
Patrick von Platen committed
1166

1167
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
1168
1169
1170
1171

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
1172
class DownBlock2D(nn.Module):
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_downsample=True,
Patrick von Platen's avatar
Patrick von Platen committed
1187
        downsample_padding=1,
1188
1189
1190
1191
1192
1193
1194
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1195
                ResnetBlock2D(
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
Patrick von Platen's avatar
Patrick von Platen committed
1213
1214
                [
                    Downsample2D(
1215
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
1216
1217
                    )
                ]
1218
1219
1220
1221
            )
        else:
            self.downsamplers = None

1222
1223
        self.gradient_checkpointing = False

1224
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
1225
1226
1227
        output_states = ()

        for resnet in self.resnets:
1228
1229
1230
1231
1232
1233
1234
1235
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

1236
1237
1238
1239
1240
1241
1242
1243
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
1244
            else:
1245
                hidden_states = resnet(hidden_states, temb, scale=scale)
1246

1247
            output_states = output_states + (hidden_states,)
1248
1249
1250

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1251
                hidden_states = downsampler(hidden_states, scale=scale)
1252

1253
            output_states = output_states + (hidden_states,)
1254
1255
1256
1257

        return hidden_states, output_states


1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
class DownEncoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1280
                ResnetBlock2D(
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=None,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1300
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1301
1302
1303
1304
1305
1306
                    )
                ]
            )
        else:
            self.downsamplers = None

1307
    def forward(self, hidden_states, scale: float = 1.0):
1308
        for resnet in self.resnets:
1309
            hidden_states = resnet(hidden_states, temb=None, scale=scale)
1310
1311
1312

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1313
                hidden_states = downsampler(hidden_states, scale)
1314
1315
1316
1317

        return hidden_states


1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
class AttnDownEncoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
1330
        attention_head_dim=1,
1331
1332
1333
1334
1335
1336
1337
1338
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []
        attentions = []

1339
1340
1341
1342
1343
1344
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

1345
1346
1347
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1348
                ResnetBlock2D(
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=None,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
1362
                Attention(
1363
                    out_channels,
1364
1365
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1366
1367
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
1368
                    norm_num_groups=resnet_groups,
1369
1370
1371
1372
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1383
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1384
1385
1386
1387
1388
1389
                    )
                ]
            )
        else:
            self.downsamplers = None

1390
    def forward(self, hidden_states, scale: float = 1.0):
1391
        for resnet, attn in zip(self.resnets, self.attentions):
1392
1393
1394
            hidden_states = resnet(hidden_states, temb=None, scale=scale)
            cross_attention_kwargs = {"scale": scale}
            hidden_states = attn(hidden_states, **cross_attention_kwargs)
1395
1396
1397

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1398
                hidden_states = downsampler(hidden_states, scale)
1399
1400
1401
1402

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
1403
class AttnSkipDownBlock2D(nn.Module):
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_pre_norm: bool = True,
1415
        attention_head_dim=1,
1416
1417
1418
1419
1420
1421
1422
        output_scale_factor=np.sqrt(2.0),
        add_downsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        self.resnets = nn.ModuleList([])

1423
1424
1425
1426
1427
1428
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

1429
1430
1431
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
1432
                ResnetBlock2D(
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(in_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            self.attentions.append(
1447
                Attention(
1448
                    out_channels,
1449
1450
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1451
1452
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
1453
1454
1455
1456
1457
                    norm_num_groups=32,
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1458
1459
1460
1461
                )
            )

        if add_downsample:
1462
            self.resnet_down = ResnetBlock2D(
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
1473
                use_in_shortcut=True,
1474
1475
1476
                down=True,
                kernel="fir",
            )
1477
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1478
1479
1480
1481
1482
1483
            self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
        else:
            self.resnet_down = None
            self.downsamplers = None
            self.skip_conv = None

1484
    def forward(self, hidden_states, temb=None, skip_sample=None, scale: float = 1.0):
1485
1486
1487
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
1488
1489
1490
            hidden_states = resnet(hidden_states, temb, scale=scale)
            cross_attention_kwargs = {"scale": scale}
            hidden_states = attn(hidden_states, **cross_attention_kwargs)
1491
1492
1493
            output_states += (hidden_states,)

        if self.downsamplers is not None:
1494
            hidden_states = self.resnet_down(hidden_states, temb, scale=scale)
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
            for downsampler in self.downsamplers:
                skip_sample = downsampler(skip_sample)

            hidden_states = self.skip_conv(skip_sample) + hidden_states

            output_states += (hidden_states,)

        return hidden_states, output_states, skip_sample


Patrick von Platen's avatar
Patrick von Platen committed
1505
class SkipDownBlock2D(nn.Module):
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_pre_norm: bool = True,
        output_scale_factor=np.sqrt(2.0),
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        self.resnets = nn.ModuleList([])

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
1527
                ResnetBlock2D(
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(in_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        if add_downsample:
1543
            self.resnet_down = ResnetBlock2D(
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
1554
                use_in_shortcut=True,
1555
1556
1557
                down=True,
                kernel="fir",
            )
1558
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1559
1560
1561
1562
1563
1564
            self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
        else:
            self.resnet_down = None
            self.downsamplers = None
            self.skip_conv = None

1565
    def forward(self, hidden_states, temb=None, skip_sample=None, scale: float = 1.0):
1566
1567
1568
        output_states = ()

        for resnet in self.resnets:
1569
            hidden_states = resnet(hidden_states, temb, scale)
1570
1571
1572
            output_states += (hidden_states,)

        if self.downsamplers is not None:
1573
            hidden_states = self.resnet_down(hidden_states, temb, scale)
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
            for downsampler in self.downsamplers:
                skip_sample = downsampler(skip_sample)

            hidden_states = self.skip_conv(skip_sample) + hidden_states

            output_states += (hidden_states,)

        return hidden_states, output_states, skip_sample


Will Berman's avatar
Will Berman committed
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
class ResnetDownsampleBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_downsample=True,
1599
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
1618
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
1638
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1639
1640
1641
1642
1643
1644
1645
1646
1647
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

1648
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
Will Berman's avatar
Will Berman committed
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
        output_states = ()

        for resnet in self.resnets:
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

1660
1661
1662
1663
1664
1665
1666
1667
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
Will Berman's avatar
Will Berman committed
1668
            else:
1669
                hidden_states = resnet(hidden_states, temb, scale)
Will Berman's avatar
Will Berman committed
1670

1671
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1672
1673
1674

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1675
                hidden_states = downsampler(hidden_states, temb, scale)
Will Berman's avatar
Will Berman committed
1676

1677
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694

        return hidden_states, output_states


class SimpleCrossAttnDownBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
1695
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
1696
1697
1698
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_downsample=True,
1699
        skip_time_act=False,
1700
        only_cross_attention=False,
1701
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
1702
1703
1704
1705
1706
1707
1708
1709
    ):
        super().__init__()

        self.has_cross_attention = True

        resnets = []
        attentions = []

1710
1711
        self.attention_head_dim = attention_head_dim
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
1727
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1728
1729
                )
            )
1730
1731
1732
1733
1734

            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
1735
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
1736
                Attention(
Will Berman's avatar
Will Berman committed
1737
1738
1739
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
1740
                    dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
1741
1742
1743
1744
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
1745
                    only_cross_attention=only_cross_attention,
1746
                    cross_attention_norm=cross_attention_norm,
1747
                    processor=processor,
Will Berman's avatar
Will Berman committed
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
                )
            )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
1767
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1768
1769
1770
1771
1772
1773
1774
1775
1776
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

1777
    def forward(
1778
1779
1780
1781
1782
1783
1784
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
1785
    ):
Will Berman's avatar
Will Berman committed
1786
        output_states = ()
1787
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
Will Berman's avatar
Will Berman committed
1788

1789
1790
        lora_scale = cross_attention_kwargs.get("scale", 1.0)

1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
        if attention_mask is None:
            # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
            mask = None if encoder_hidden_states is None else encoder_attention_mask
        else:
            # when attention_mask is defined: we don't even check for encoder_attention_mask.
            # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
            # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
            #       then we can simplify this whole if/else block to:
            #         mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
            mask = attention_mask

Will Berman's avatar
Will Berman committed
1802
        for resnet, attn in zip(self.resnets, self.attentions):
1803
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
1804

1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
ethansmith2000's avatar
ethansmith2000 committed
1815
                hidden_states = attn(
1816
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
1817
1818
1819
1820
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=mask,
                    **cross_attention_kwargs,
                )
1821
            else:
1822
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
1823
1824
1825
1826

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
1827
                    attention_mask=mask,
1828
1829
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
1830

1831
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1832
1833
1834

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1835
                hidden_states = downsampler(hidden_states, temb, scale=lora_scale)
Will Berman's avatar
Will Berman committed
1836

1837
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1838
1839
1840
1841

        return hidden_states, output_states


1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
class KDownBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 4,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
        resnet_group_size: int = 32,
        add_downsample=False,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            groups = in_channels // resnet_group_size
            groups_out = out_channels // resnet_group_size

            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    dropout=dropout,
                    temb_channels=temb_channels,
                    groups=groups,
                    groups_out=groups_out,
                    eps=resnet_eps,
                    non_linearity=resnet_act_fn,
                    time_embedding_norm="ada_group",
                    conv_shortcut_bias=False,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            # YiYi's comments- might be able to use FirDownsample2D, look into details later
            self.downsamplers = nn.ModuleList([KDownsample2D()])
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

1888
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
        output_states = ()

        for resnet in self.resnets:
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

1900
1901
1902
1903
1904
1905
1906
1907
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
1908
            else:
1909
                hidden_states = resnet(hidden_states, temb, scale)
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930

            output_states += (hidden_states,)

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
                hidden_states = downsampler(hidden_states)

        return hidden_states, output_states


class KCrossAttnDownBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        cross_attention_dim: int,
        dropout: float = 0.0,
        num_layers: int = 4,
        resnet_group_size: int = 32,
        add_downsample=True,
1931
        attention_head_dim: int = 64,
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
        add_self_attention: bool = False,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            groups = in_channels // resnet_group_size
            groups_out = out_channels // resnet_group_size

            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    dropout=dropout,
                    temb_channels=temb_channels,
                    groups=groups,
                    groups_out=groups_out,
                    eps=resnet_eps,
                    non_linearity=resnet_act_fn,
                    time_embedding_norm="ada_group",
                    conv_shortcut_bias=False,
                )
            )
            attentions.append(
                KAttentionBlock(
                    out_channels,
1964
1965
                    out_channels // attention_head_dim,
                    attention_head_dim,
1966
1967
1968
1969
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
1970
                    cross_attention_norm="layer_norm",
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
                    group_size=resnet_group_size,
                )
            )

        self.resnets = nn.ModuleList(resnets)
        self.attentions = nn.ModuleList(attentions)

        if add_downsample:
            self.downsamplers = nn.ModuleList([KDownsample2D()])
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(
1986
1987
1988
1989
1990
1991
1992
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
1993
1994
    ):
        output_states = ()
1995
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008

        for resnet, attn in zip(self.resnets, self.attentions):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

2009
2010
2011
2012
2013
2014
2015
                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
ethansmith2000's avatar
ethansmith2000 committed
2016
                hidden_states = attn(
2017
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
2018
2019
2020
2021
2022
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
2023
                )
2024
            else:
2025
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
2026
2027
2028
2029
2030
2031
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
2032
                    encoder_attention_mask=encoder_attention_mask,
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
                )

            if self.downsamplers is None:
                output_states += (None,)
            else:
                output_states += (hidden_states,)

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
                hidden_states = downsampler(hidden_states)

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
2047
class AttnUpBlock2D(nn.Module):
2048
2049
2050
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
2051
2052
        prev_output_channel: int,
        out_channels: int,
2053
        temb_channels: int,
2054
        resolution_idx: int = None,
2055
2056
2057
2058
2059
2060
2061
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
2062
        attention_head_dim=1,
2063
        output_scale_factor=1.0,
2064
        upsample_type="conv",
2065
2066
2067
2068
2069
    ):
        super().__init__()
        resnets = []
        attentions = []

2070
2071
        self.upsample_type = upsample_type

2072
2073
2074
2075
2076
2077
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

2078
        for i in range(num_layers):
Patrick von Platen's avatar
Patrick von Platen committed
2079
2080
2081
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

2082
            resnets.append(
2083
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2084
2085
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
2097
                Attention(
Patrick von Platen's avatar
Patrick von Platen committed
2098
                    out_channels,
2099
2100
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
2101
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
2102
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
2103
                    norm_num_groups=resnet_groups,
2104
2105
2106
2107
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
2108
2109
2110
2111
2112
2113
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

2114
        if upsample_type == "conv":
Patrick von Platen's avatar
Patrick von Platen committed
2115
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
        elif upsample_type == "resnet":
            self.upsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
                        up=True,
                    )
                ]
            )
2134
2135
2136
        else:
            self.upsamplers = None

2137
2138
        self.resolution_idx = resolution_idx

2139
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, scale: float = 1.0):
2140
2141
2142
2143
2144
2145
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

2146
2147
2148
            hidden_states = resnet(hidden_states, temb, scale=scale)
            cross_attention_kwargs = {"scale": scale}
            hidden_states = attn(hidden_states, **cross_attention_kwargs)
2149
2150
2151

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2152
                if self.upsample_type == "resnet":
2153
                    hidden_states = upsampler(hidden_states, temb=temb, scale=scale)
2154
                else:
2155
                    hidden_states = upsampler(hidden_states, scale=scale)
2156
2157
2158
2159

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2160
class CrossAttnUpBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
2161
2162
2163
2164
2165
2166
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
2167
        resolution_idx: int = None,
Patrick von Platen's avatar
Patrick von Platen committed
2168
2169
        dropout: float = 0.0,
        num_layers: int = 1,
2170
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
2171
2172
2173
2174
2175
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
2176
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
2177
2178
2179
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2180
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
2181
        use_linear_projection=False,
2182
        only_cross_attention=False,
2183
        upcast_attention=False,
2184
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
2185
2186
2187
2188
2189
    ):
        super().__init__()
        resnets = []
        attentions = []

2190
        self.has_cross_attention = True
2191
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
2192
2193
2194
2195
2196
2197

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            resnets.append(
2198
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
2211
2212
2213
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
2214
2215
                        num_attention_heads,
                        out_channels // num_attention_heads,
2216
                        in_channels=out_channels,
2217
                        num_layers=transformer_layers_per_block,
2218
2219
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
2220
                        use_linear_projection=use_linear_projection,
2221
                        only_cross_attention=only_cross_attention,
2222
                        upcast_attention=upcast_attention,
2223
                        attention_type=attention_type,
2224
2225
2226
2227
2228
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
2229
2230
                        num_attention_heads,
                        out_channels // num_attention_heads,
2231
2232
2233
2234
2235
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
Patrick von Platen's avatar
Patrick von Platen committed
2236
2237
2238
2239
2240
2241
2242
2243
2244
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

2245
        self.gradient_checkpointing = False
2246
        self.resolution_idx = resolution_idx
2247
2248
2249

    def forward(
        self,
2250
2251
2252
2253
2254
2255
2256
2257
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
2258
    ):
2259
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
2260
2261
2262
2263
2264
2265
        is_freeu_enabled = (
            getattr(self, "s1", None)
            and getattr(self, "s2", None)
            and getattr(self, "b1", None)
            and getattr(self, "b2", None)
        )
2266

Patrick von Platen's avatar
Patrick von Platen committed
2267
2268
2269
2270
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283

            # FreeU: Only operate on the first two stages
            if is_freeu_enabled:
                hidden_states, res_hidden_states = apply_freeu(
                    self.resolution_idx,
                    hidden_states,
                    res_hidden_states,
                    s1=self.s1,
                    s2=self.s2,
                    b1=self.b1,
                    b2=self.b2,
                )

Patrick von Platen's avatar
Patrick von Platen committed
2284
2285
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

2286
2287
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
2288
                def create_custom_forward(module, return_dict=None):
2289
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
2290
2291
2292
2293
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
2294
2295
2296

                    return custom_forward

2297
2298
2299
2300
2301
2302
2303
                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
ethansmith2000's avatar
ethansmith2000 committed
2304
                hidden_states = attn(
2305
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
2306
2307
2308
2309
2310
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
2311
                )[0]
2312
            else:
2313
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
2314
2315
2316
2317
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
2318
2319
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
2320
2321
                    return_dict=False,
                )[0]
Patrick von Platen's avatar
Patrick von Platen committed
2322
2323
2324

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2325
                hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
Patrick von Platen's avatar
Patrick von Platen committed
2326
2327
2328
2329

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2330
class UpBlock2D(nn.Module):
2331
2332
2333
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
2334
2335
        prev_output_channel: int,
        out_channels: int,
2336
        temb_channels: int,
2337
        resolution_idx: int = None,
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
Patrick von Platen's avatar
Patrick von Platen committed
2352
2353
2354
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

2355
            resnets.append(
2356
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2357
2358
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
Patrick von Platen's avatar
Patrick von Platen committed
2373
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
2374
2375
2376
        else:
            self.upsamplers = None

2377
        self.gradient_checkpointing = False
2378
        self.resolution_idx = resolution_idx
2379

2380
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, scale: float = 1.0):
2381
2382
2383
2384
2385
2386
2387
        is_freeu_enabled = (
            getattr(self, "s1", None)
            and getattr(self, "s2", None)
            and getattr(self, "b1", None)
            and getattr(self, "b2", None)
        )

2388
2389
2390
2391
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404

            # FreeU: Only operate on the first two stages
            if is_freeu_enabled:
                hidden_states, res_hidden_states = apply_freeu(
                    self.resolution_idx,
                    hidden_states,
                    res_hidden_states,
                    s1=self.s1,
                    s2=self.s2,
                    b1=self.b1,
                    b2=self.b2,
                )

2405
2406
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

2407
2408
2409
2410
2411
2412
2413
2414
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

2415
2416
2417
2418
2419
2420
2421
2422
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
2423
            else:
2424
                hidden_states = resnet(hidden_states, temb, scale=scale)
2425
2426
2427

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2428
                hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
2429
2430

        return hidden_states
2431
2432


2433
2434
2435
2436
2437
class UpDecoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
2438
        resolution_idx: int = None,
2439
2440
2441
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
YiYi Xu's avatar
YiYi Xu committed
2442
        resnet_time_scale_shift: str = "default",  # default, spatial
2443
2444
2445
2446
2447
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
YiYi Xu's avatar
YiYi Xu committed
2448
        temb_channels=None,
2449
2450
2451
2452
2453
2454
2455
2456
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
2457
                ResnetBlock2D(
2458
2459
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2460
                    temb_channels=temb_channels,
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

2478
2479
        self.resolution_idx = resolution_idx

2480
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
2481
        for resnet in self.resnets:
2482
            hidden_states = resnet(hidden_states, temb=temb, scale=scale)
2483
2484
2485
2486
2487
2488
2489
2490

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states)

        return hidden_states


2491
2492
2493
2494
2495
class AttnUpDecoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
2496
        resolution_idx: int = None,
2497
2498
2499
2500
2501
2502
2503
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
2504
        attention_head_dim=1,
2505
2506
        output_scale_factor=1.0,
        add_upsample=True,
YiYi Xu's avatar
YiYi Xu committed
2507
        temb_channels=None,
2508
2509
2510
2511
2512
    ):
        super().__init__()
        resnets = []
        attentions = []

2513
2514
2515
2516
2517
2518
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

2519
2520
2521
2522
        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
2523
                ResnetBlock2D(
2524
2525
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2526
                    temb_channels=temb_channels,
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
2537
                Attention(
2538
                    out_channels,
2539
2540
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
2541
2542
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
2543
                    norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
YiYi Xu's avatar
YiYi Xu committed
2544
                    spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
2545
2546
2547
2548
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

2560
2561
        self.resolution_idx = resolution_idx

2562
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
2563
        for resnet, attn in zip(self.resnets, self.attentions):
2564
2565
2566
            hidden_states = resnet(hidden_states, temb=temb, scale=scale)
            cross_attention_kwargs = {"scale": scale}
            hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs)
2567
2568
2569

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2570
                hidden_states = upsampler(hidden_states, scale=scale)
2571
2572
2573
2574

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2575
class AttnSkipUpBlock2D(nn.Module):
2576
2577
2578
2579
2580
2581
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
2582
        resolution_idx: int = None,
2583
2584
2585
2586
2587
2588
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_pre_norm: bool = True,
2589
        attention_head_dim=1,
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
        output_scale_factor=np.sqrt(2.0),
        add_upsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        self.resnets = nn.ModuleList([])

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            self.resnets.append(
2602
                ResnetBlock2D(
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(resnet_in_channels + res_skip_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

2617
2618
2619
2620
2621
2622
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

2623
        self.attentions.append(
2624
            Attention(
2625
                out_channels,
2626
2627
                heads=out_channels // attention_head_dim,
                dim_head=attention_head_dim,
2628
2629
                rescale_output_factor=output_scale_factor,
                eps=resnet_eps,
2630
2631
2632
2633
2634
                norm_num_groups=32,
                residual_connection=True,
                bias=True,
                upcast_softmax=True,
                _from_deprecated_attn_block=True,
2635
2636
2637
2638
2639
            )
        )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
2640
            self.resnet_up = ResnetBlock2D(
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                groups_out=min(out_channels // 4, 32),
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
2652
                use_in_shortcut=True,
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
                up=True,
                kernel="fir",
            )
            self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
            self.skip_norm = torch.nn.GroupNorm(
                num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
            )
            self.act = nn.SiLU()
        else:
            self.resnet_up = None
            self.skip_conv = None
            self.skip_norm = None
            self.act = None

2667
2668
        self.resolution_idx = resolution_idx

2669
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None, scale: float = 1.0):
2670
2671
2672
2673
2674
2675
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

2676
            hidden_states = resnet(hidden_states, temb, scale=scale)
2677

2678
2679
        cross_attention_kwargs = {"scale": scale}
        hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs)
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692

        if skip_sample is not None:
            skip_sample = self.upsampler(skip_sample)
        else:
            skip_sample = 0

        if self.resnet_up is not None:
            skip_sample_states = self.skip_norm(hidden_states)
            skip_sample_states = self.act(skip_sample_states)
            skip_sample_states = self.skip_conv(skip_sample_states)

            skip_sample = skip_sample + skip_sample_states

2693
            hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
2694
2695
2696
2697

        return hidden_states, skip_sample


Patrick von Platen's avatar
Patrick von Platen committed
2698
class SkipUpBlock2D(nn.Module):
2699
2700
2701
2702
2703
2704
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
2705
        resolution_idx: int = None,
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_pre_norm: bool = True,
        output_scale_factor=np.sqrt(2.0),
        add_upsample=True,
        upsample_padding=1,
    ):
        super().__init__()
        self.resnets = nn.ModuleList([])

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            self.resnets.append(
2724
                ResnetBlock2D(
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
2741
            self.resnet_up = ResnetBlock2D(
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                groups_out=min(out_channels // 4, 32),
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
2753
                use_in_shortcut=True,
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
                up=True,
                kernel="fir",
            )
            self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
            self.skip_norm = torch.nn.GroupNorm(
                num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
            )
            self.act = nn.SiLU()
        else:
            self.resnet_up = None
            self.skip_conv = None
            self.skip_norm = None
            self.act = None

2768
2769
        self.resolution_idx = resolution_idx

2770
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None, scale: float = 1.0):
2771
2772
2773
2774
2775
2776
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

2777
            hidden_states = resnet(hidden_states, temb, scale=scale)
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790

        if skip_sample is not None:
            skip_sample = self.upsampler(skip_sample)
        else:
            skip_sample = 0

        if self.resnet_up is not None:
            skip_sample_states = self.skip_norm(hidden_states)
            skip_sample_states = self.act(skip_sample_states)
            skip_sample_states = self.skip_conv(skip_sample_states)

            skip_sample = skip_sample + skip_sample_states

2791
            hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
2792
2793

        return hidden_states, skip_sample
Will Berman's avatar
Will Berman committed
2794
2795
2796
2797
2798
2799
2800
2801
2802


class ResnetUpsampleBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
2803
        resolution_idx: int = None,
Will Berman's avatar
Will Berman committed
2804
2805
2806
2807
2808
2809
2810
2811
2812
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
2813
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
2834
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
2854
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2855
2856
2857
2858
2859
2860
2861
2862
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False
2863
        self.resolution_idx = resolution_idx
Will Berman's avatar
Will Berman committed
2864

2865
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, scale: float = 1.0):
Will Berman's avatar
Will Berman committed
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

2880
2881
2882
2883
2884
2885
2886
2887
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
Will Berman's avatar
Will Berman committed
2888
            else:
2889
                hidden_states = resnet(hidden_states, temb, scale=scale)
Will Berman's avatar
Will Berman committed
2890
2891
2892

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2893
                hidden_states = upsampler(hidden_states, temb, scale=scale)
Will Berman's avatar
Will Berman committed
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904

        return hidden_states


class SimpleCrossAttnUpBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
2905
        resolution_idx: int = None,
Will Berman's avatar
Will Berman committed
2906
2907
2908
2909
2910
2911
2912
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
2913
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
2914
2915
2916
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2917
        skip_time_act=False,
2918
        only_cross_attention=False,
2919
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
2920
2921
2922
2923
2924
2925
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
2926
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
2927

2928
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
2946
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2947
2948
                )
            )
2949
2950
2951
2952
2953

            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
2954
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
2955
                Attention(
Will Berman's avatar
Will Berman committed
2956
2957
2958
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
2959
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
2960
2961
2962
2963
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
2964
                    only_cross_attention=only_cross_attention,
2965
                    cross_attention_norm=cross_attention_norm,
2966
                    processor=processor,
Will Berman's avatar
Will Berman committed
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
                )
            )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_channels,
                        out_channels=out_channels,
                        temb_channels=temb_channels,
                        eps=resnet_eps,
                        groups=resnet_groups,
                        dropout=dropout,
                        time_embedding_norm=resnet_time_scale_shift,
                        non_linearity=resnet_act_fn,
                        output_scale_factor=output_scale_factor,
                        pre_norm=resnet_pre_norm,
2986
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2987
2988
2989
2990
2991
2992
2993
2994
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False
2995
        self.resolution_idx = resolution_idx
Will Berman's avatar
Will Berman committed
2996
2997
2998

    def forward(
        self,
2999
3000
3001
3002
3003
3004
3005
3006
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
Will Berman's avatar
Will Berman committed
3007
    ):
3008
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
3009

3010
        lora_scale = cross_attention_kwargs.get("scale", 1.0)
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
        if attention_mask is None:
            # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
            mask = None if encoder_hidden_states is None else encoder_attention_mask
        else:
            # when attention_mask is defined: we don't even check for encoder_attention_mask.
            # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
            # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
            #       then we can simplify this whole if/else block to:
            #         mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
            mask = attention_mask

Will Berman's avatar
Will Berman committed
3022
3023
3024
3025
3026
3027
3028
        for resnet, attn in zip(self.resnets, self.attentions):
            # resnet
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

3029
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
3030

3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
ethansmith2000's avatar
ethansmith2000 committed
3041
                hidden_states = attn(
3042
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
3043
3044
3045
3046
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=mask,
                    **cross_attention_kwargs,
                )
3047
            else:
3048
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
3049
3050
3051
3052

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
3053
                    attention_mask=mask,
3054
3055
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
3056
3057
3058

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
3059
                hidden_states = upsampler(hidden_states, temb, scale=lora_scale)
Will Berman's avatar
Will Berman committed
3060
3061

        return hidden_states
3062
3063
3064
3065
3066
3067
3068
3069


class KUpBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
3070
        resolution_idx: int,
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
        dropout: float = 0.0,
        num_layers: int = 5,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
        resnet_group_size: Optional[int] = 32,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        k_in_channels = 2 * out_channels
        k_out_channels = in_channels
        num_layers = num_layers - 1

        for i in range(num_layers):
            in_channels = k_in_channels if i == 0 else out_channels
            groups = in_channels // resnet_group_size
            groups_out = out_channels // resnet_group_size

            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=k_out_channels if (i == num_layers - 1) else out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=groups,
                    groups_out=groups_out,
                    dropout=dropout,
                    non_linearity=resnet_act_fn,
                    time_embedding_norm="ada_group",
                    conv_shortcut_bias=False,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([KUpsample2D()])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False
3112
        self.resolution_idx = resolution_idx
3113

3114
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, scale: float = 1.0):
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
        res_hidden_states_tuple = res_hidden_states_tuple[-1]
        if res_hidden_states_tuple is not None:
            hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)

        for resnet in self.resnets:
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

3128
3129
3130
3131
3132
3133
3134
3135
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
3136
            else:
3137
                hidden_states = resnet(hidden_states, temb, scale=scale)
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states)

        return hidden_states


class KCrossAttnUpBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
3152
        resolution_idx: int,
3153
3154
3155
3156
3157
        dropout: float = 0.0,
        num_layers: int = 4,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
        resnet_group_size: int = 32,
3158
        attention_head_dim=1,  # attention dim_head
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
        cross_attention_dim: int = 768,
        add_upsample: bool = True,
        upcast_attention: bool = False,
    ):
        super().__init__()
        resnets = []
        attentions = []

        is_first_block = in_channels == out_channels == temb_channels
        is_middle_block = in_channels != out_channels
        add_self_attention = True if is_first_block else False

        self.has_cross_attention = True
3172
        self.attention_head_dim = attention_head_dim
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207

        # in_channels, and out_channels for the block (k-unet)
        k_in_channels = out_channels if is_first_block else 2 * out_channels
        k_out_channels = in_channels

        num_layers = num_layers - 1

        for i in range(num_layers):
            in_channels = k_in_channels if i == 0 else out_channels
            groups = in_channels // resnet_group_size
            groups_out = out_channels // resnet_group_size

            if is_middle_block and (i == num_layers - 1):
                conv_2d_out_channels = k_out_channels
            else:
                conv_2d_out_channels = None

            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    conv_2d_out_channels=conv_2d_out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=groups,
                    groups_out=groups_out,
                    dropout=dropout,
                    non_linearity=resnet_act_fn,
                    time_embedding_norm="ada_group",
                    conv_shortcut_bias=False,
                )
            )
            attentions.append(
                KAttentionBlock(
                    k_out_channels if (i == num_layers - 1) else out_channels,
3208
                    k_out_channels // attention_head_dim
3209
                    if (i == num_layers - 1)
3210
3211
                    else out_channels // attention_head_dim,
                    attention_head_dim,
3212
3213
3214
3215
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
3216
                    cross_attention_norm="layer_norm",
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
                    upcast_attention=upcast_attention,
                )
            )

        self.resnets = nn.ModuleList(resnets)
        self.attentions = nn.ModuleList(attentions)

        if add_upsample:
            self.upsamplers = nn.ModuleList([KUpsample2D()])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False
3230
        self.resolution_idx = resolution_idx
3231
3232
3233

    def forward(
        self,
3234
3235
3236
3237
3238
3239
3240
3241
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
3242
3243
3244
3245
3246
    ):
        res_hidden_states_tuple = res_hidden_states_tuple[-1]
        if res_hidden_states_tuple is not None:
            hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)

3247
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
        for resnet, attn in zip(self.resnets, self.attentions):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

3260
3261
3262
3263
3264
3265
3266
                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
ethansmith2000's avatar
ethansmith2000 committed
3267
                hidden_states = attn(
3268
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
3269
3270
3271
3272
3273
3274
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
                )
3275
            else:
3276
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
3277
3278
3279
3280
3281
3282
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
3283
                    encoder_attention_mask=encoder_attention_mask,
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
                )

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states)

        return hidden_states


# can potentially later be renamed to `No-feed-forward` attention
class KAttentionBlock(nn.Module):
    r"""
    A basic Transformer block.

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

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout: float = 0.0,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        upcast_attention: bool = False,
        temb_channels: int = 768,  # for ada_group_norm
        add_self_attention: bool = False,
3322
        cross_attention_norm: Optional[str] = None,
3323
3324
3325
3326
3327
3328
3329
3330
        group_size: int = 32,
    ):
        super().__init__()
        self.add_self_attention = add_self_attention

        # 1. Self-Attn
        if add_self_attention:
            self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
Patrick von Platen's avatar
Patrick von Platen committed
3331
            self.attn1 = Attention(
3332
3333
3334
3335
3336
3337
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                cross_attention_dim=None,
3338
                cross_attention_norm=None,
3339
3340
3341
3342
            )

        # 2. Cross-Attn
        self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
Patrick von Platen's avatar
Patrick von Platen committed
3343
        self.attn2 = Attention(
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
            query_dim=dim,
            cross_attention_dim=cross_attention_dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            upcast_attention=upcast_attention,
            cross_attention_norm=cross_attention_norm,
        )

    def _to_3d(self, hidden_states, height, weight):
        return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1)

    def _to_4d(self, hidden_states, height, weight):
        return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight)

    def forward(
        self,
3362
3363
3364
3365
3366
3367
3368
3369
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        # TODO: mark emb as non-optional (self.norm2 requires it).
        #       requires assessing impact of change to positional param interface.
        emb: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}

        # 1. Self-Attention
        if self.add_self_attention:
            norm_hidden_states = self.norm1(hidden_states, emb)

            height, weight = norm_hidden_states.shape[2:]
            norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)

            attn_output = self.attn1(
                norm_hidden_states,
                encoder_hidden_states=None,
3383
                attention_mask=attention_mask,
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
                **cross_attention_kwargs,
            )
            attn_output = self._to_4d(attn_output, height, weight)

            hidden_states = attn_output + hidden_states

        # 2. Cross-Attention/None
        norm_hidden_states = self.norm2(hidden_states, emb)

        height, weight = norm_hidden_states.shape[2:]
        norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
        attn_output = self.attn2(
            norm_hidden_states,
            encoder_hidden_states=encoder_hidden_states,
3398
            attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
3399
3400
3401
3402
3403
3404
3405
            **cross_attention_kwargs,
        )
        attn_output = self._to_4d(attn_output, height, weight)

        hidden_states = attn_output + hidden_states

        return hidden_states