unet_2d_blocks.py 125 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
470
471
472
473
474
475
476
477
478
479
480
    """
    Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU blocks.

    Args:
        in_channels (`int`): The number of input channels.
        out_channels (`int`): The number of output channels.
        act_fn (`str`):` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.

    Returns:
        `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to `out_channels`.
    """

481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
    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
502
503
504
505
506
class UNetMidBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
507
        dropout: float = 0.0,
508
        num_layers: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
509
        resnet_eps: float = 1e-6,
YiYi Xu's avatar
YiYi Xu committed
510
        resnet_time_scale_shift: str = "default",  # default, spatial
Patrick von Platen's avatar
Patrick von Platen committed
511
512
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
513
        attn_groups: Optional[int] = None,
514
        resnet_pre_norm: bool = True,
Will Berman's avatar
Will Berman committed
515
        add_attention: bool = True,
516
        attention_head_dim=1,
Patrick von Platen's avatar
Patrick von Platen committed
517
518
519
        output_scale_factor=1.0,
    ):
        super().__init__()
520
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
Will Berman's avatar
Will Berman committed
521
        self.add_attention = add_attention
Patrick von Platen's avatar
Patrick von Platen committed
522

523
524
525
        if attn_groups is None:
            attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None

526
527
        # there is always at least one resnet
        resnets = [
528
            ResnetBlock2D(
529
530
531
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
532
                eps=resnet_eps,
533
534
535
536
537
538
                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
539
            )
540
541
        ]
        attentions = []
Patrick von Platen's avatar
Patrick von Platen committed
542

543
544
545
546
547
548
        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

549
        for _ in range(num_layers):
Will Berman's avatar
Will Berman committed
550
551
            if self.add_attention:
                attentions.append(
552
                    Attention(
Will Berman's avatar
Will Berman committed
553
                        in_channels,
554
555
                        heads=in_channels // attention_head_dim,
                        dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
556
557
                        rescale_output_factor=output_scale_factor,
                        eps=resnet_eps,
558
                        norm_num_groups=attn_groups,
YiYi Xu's avatar
YiYi Xu committed
559
                        spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
560
561
562
563
                        residual_connection=True,
                        bias=True,
                        upcast_softmax=True,
                        _from_deprecated_attn_block=True,
Will Berman's avatar
Will Berman committed
564
                    )
565
                )
Will Berman's avatar
Will Berman committed
566
567
568
            else:
                attentions.append(None)

569
            resnets.append(
570
                ResnetBlock2D(
571
572
573
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
574
                    eps=resnet_eps,
575
576
577
578
579
580
581
                    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
582
583
            )

584
585
586
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

Will Berman's avatar
Will Berman committed
587
    def forward(self, hidden_states, temb=None):
Patrick von Platen's avatar
Patrick von Platen committed
588
        hidden_states = self.resnets[0](hidden_states, temb)
589
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
Will Berman's avatar
Will Berman committed
590
            if attn is not None:
YiYi Xu's avatar
YiYi Xu committed
591
                hidden_states = attn(hidden_states, temb=temb)
Patrick von Platen's avatar
Patrick von Platen committed
592
            hidden_states = resnet(hidden_states, temb)
Patrick von Platen's avatar
Patrick von Platen committed
593

594
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
595

596

Patrick von Platen's avatar
Patrick von Platen committed
597
598
599
600
601
602
603
class UNetMidBlock2DCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
604
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
605
606
607
608
609
        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,
610
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
611
612
        output_scale_factor=1.0,
        cross_attention_dim=1280,
613
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
614
        use_linear_projection=False,
615
        upcast_attention=False,
616
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
617
618
619
    ):
        super().__init__()

620
        self.has_cross_attention = True
621
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
622
623
624
625
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

        # there is always at least one resnet
        resnets = [
626
            ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
                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):
642
643
644
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
645
646
                        num_attention_heads,
                        in_channels // num_attention_heads,
647
                        in_channels=in_channels,
648
                        num_layers=transformer_layers_per_block,
649
650
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
651
                        use_linear_projection=use_linear_projection,
652
                        upcast_attention=upcast_attention,
653
                        attention_type=attention_type,
654
655
656
657
658
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
659
660
                        num_attention_heads,
                        in_channels // num_attention_heads,
661
662
663
664
665
                        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
666
667
                )
            resnets.append(
668
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
                    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)

685
686
        self.gradient_checkpointing = False

687
    def forward(
688
689
690
691
692
693
694
695
        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:
696
697
        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
698
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
699
700
701
702
703
704
705
706
707
708
709
710
            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
711
                hidden_states = attn(
712
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
713
714
715
716
717
                    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,
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
                )[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]
734
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
Will Berman's avatar
Will Berman committed
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750

        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,
751
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
752
753
        output_scale_factor=1.0,
        cross_attention_dim=1280,
754
        skip_time_act=False,
755
        only_cross_attention=False,
756
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
757
758
759
760
761
    ):
        super().__init__()

        self.has_cross_attention = True

762
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
763
764
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

765
        self.num_heads = in_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
766
767
768
769
770
771
772
773
774
775
776
777
778
779

        # 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,
780
                skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
781
782
783
784
785
            )
        ]
        attentions = []

        for _ in range(num_layers):
786
787
788
789
            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
790
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
791
                Attention(
Will Berman's avatar
Will Berman committed
792
793
794
                    query_dim=in_channels,
                    cross_attention_dim=in_channels,
                    heads=self.num_heads,
795
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
796
797
798
799
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
800
                    only_cross_attention=only_cross_attention,
801
                    cross_attention_norm=cross_attention_norm,
802
                    processor=processor,
Will Berman's avatar
Will Berman committed
803
804
805
806
807
808
809
810
811
812
813
814
815
816
                )
            )
            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,
817
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
818
819
820
821
822
823
                )
            )

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

824
    def forward(
825
826
827
828
829
830
831
        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,
832
833
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
834
        lora_scale = cross_attention_kwargs.get("scale", 1.0)
835
836
837
838
839
840
841
842
843
844
845
846

        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

847
        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
Will Berman's avatar
Will Berman committed
848
849
850
851
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            # attn
            hidden_states = attn(
                hidden_states,
852
                encoder_hidden_states=encoder_hidden_states,
853
                attention_mask=mask,
854
                **cross_attention_kwargs,
Will Berman's avatar
Will Berman committed
855
856
857
            )

            # resnet
858
            hidden_states = resnet(hidden_states, temb, scale=lora_scale)
Patrick von Platen's avatar
Patrick von Platen committed
859
860
861
862

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
863
class AttnDownBlock2D(nn.Module):
864
865
866
867
868
869
870
871
872
873
874
875
    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,
876
        attention_head_dim=1,
877
        output_scale_factor=1.0,
878
        downsample_padding=1,
879
        downsample_type="conv",
880
881
882
883
    ):
        super().__init__()
        resnets = []
        attentions = []
884
        self.downsample_type = downsample_type
885

886
887
888
889
890
891
        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

892
893
894
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
895
                ResnetBlock2D(
896
897
898
899
900
901
902
903
904
905
906
907
908
                    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(
909
                Attention(
910
                    out_channels,
911
912
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
913
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
914
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
915
                    norm_num_groups=resnet_groups,
916
917
918
919
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
920
921
922
923
924
925
                )
            )

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

926
        if downsample_type == "conv":
927
            self.downsamplers = nn.ModuleList(
928
929
                [
                    Downsample2D(
930
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
931
932
                    )
                ]
933
            )
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
        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,
                    )
                ]
            )
952
953
954
        else:
            self.downsamplers = None

955
956
957
958
959
    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)

960
961
962
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
963
964
965
            cross_attention_kwargs.update({"scale": lora_scale})
            hidden_states = resnet(hidden_states, temb, scale=lora_scale)
            hidden_states = attn(hidden_states, **cross_attention_kwargs)
966
            output_states = output_states + (hidden_states,)
967
968
969

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
970
                if self.downsample_type == "resnet":
971
                    hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale)
972
                else:
973
                    hidden_states = downsampler(hidden_states, scale=lora_scale)
974
975
976
977
978
979

            output_states += (hidden_states,)

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
980
class CrossAttnDownBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
981
982
983
984
985
986
987
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
988
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
989
990
991
992
993
        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,
994
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
995
996
997
998
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
999
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
1000
        use_linear_projection=False,
1001
        only_cross_attention=False,
1002
        upcast_attention=False,
1003
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
1004
1005
1006
1007
1008
    ):
        super().__init__()
        resnets = []
        attentions = []

1009
        self.has_cross_attention = True
1010
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
1011
1012
1013
1014

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1015
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
                    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,
                )
            )
1028
1029
1030
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
1031
1032
                        num_attention_heads,
                        out_channels // num_attention_heads,
1033
                        in_channels=out_channels,
1034
                        num_layers=transformer_layers_per_block,
1035
1036
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
1037
                        use_linear_projection=use_linear_projection,
1038
                        only_cross_attention=only_cross_attention,
1039
                        upcast_attention=upcast_attention,
1040
                        attention_type=attention_type,
1041
1042
1043
1044
1045
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
1046
1047
                        num_attention_heads,
                        out_channels // num_attention_heads,
1048
1049
1050
1051
1052
                        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
1053
1054
1055
1056
1057
1058
1059
1060
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1061
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
1062
1063
1064
1065
1066
1067
                    )
                ]
            )
        else:
            self.downsamplers = None

1068
1069
        self.gradient_checkpointing = False

1070
    def forward(
1071
1072
1073
1074
1075
1076
1077
        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
1078
        additional_residuals=None,
1079
    ):
Patrick von Platen's avatar
Patrick von Platen committed
1080
1081
        output_states = ()

1082
1083
        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
1084
1085
1086
        blocks = list(zip(self.resnets, self.attentions))

        for i, (resnet, attn) in enumerate(blocks):
1087
1088
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
1089
                def create_custom_forward(module, return_dict=None):
1090
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
1091
1092
1093
1094
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
1095
1096
1097

                    return custom_forward

1098
1099
1100
1101
1102
1103
1104
                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
1105
                hidden_states = attn(
1106
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
1107
1108
1109
1110
1111
                    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,
1112
                )[0]
1113
            else:
1114
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
1115
1116
1117
1118
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
1119
1120
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
1121
1122
                    return_dict=False,
                )[0]
1123

Will Berman's avatar
Will Berman committed
1124
1125
1126
1127
            # 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

1128
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
1129
1130
1131

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1132
                hidden_states = downsampler(hidden_states, scale=lora_scale)
Patrick von Platen's avatar
Patrick von Platen committed
1133

1134
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
1135
1136
1137
1138

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
1139
class DownBlock2D(nn.Module):
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
    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
1154
        downsample_padding=1,
1155
1156
1157
1158
1159
1160
1161
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1162
                ResnetBlock2D(
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
                    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
1180
1181
                [
                    Downsample2D(
1182
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
1183
1184
                    )
                ]
1185
1186
1187
1188
            )
        else:
            self.downsamplers = None

1189
1190
        self.gradient_checkpointing = False

1191
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
1192
1193
1194
        output_states = ()

        for resnet in self.resnets:
1195
1196
1197
1198
1199
1200
1201
1202
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

1203
1204
1205
1206
1207
1208
1209
1210
                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
                    )
1211
            else:
1212
                hidden_states = resnet(hidden_states, temb, scale=scale)
1213

1214
            output_states = output_states + (hidden_states,)
1215
1216
1217

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1218
                hidden_states = downsampler(hidden_states, scale=scale)
1219

1220
            output_states = output_states + (hidden_states,)
1221
1222
1223
1224

        return hidden_states, output_states


1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
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(
1247
                ResnetBlock2D(
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
                    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(
1267
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1268
1269
1270
1271
1272
1273
                    )
                ]
            )
        else:
            self.downsamplers = None

1274
    def forward(self, hidden_states, scale: float = 1.0):
1275
        for resnet in self.resnets:
1276
            hidden_states = resnet(hidden_states, temb=None, scale=scale)
1277
1278
1279

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1280
                hidden_states = downsampler(hidden_states, scale)
1281
1282
1283
1284

        return hidden_states


1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
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,
1297
        attention_head_dim=1,
1298
1299
1300
1301
1302
1303
1304
1305
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []
        attentions = []

1306
1307
1308
1309
1310
1311
        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

1312
1313
1314
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1315
                ResnetBlock2D(
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
                    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(
1329
                Attention(
1330
                    out_channels,
1331
1332
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1333
1334
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
1335
                    norm_num_groups=resnet_groups,
1336
1337
1338
1339
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
                )
            )

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

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1350
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1351
1352
1353
1354
1355
1356
                    )
                ]
            )
        else:
            self.downsamplers = None

1357
    def forward(self, hidden_states, scale: float = 1.0):
1358
        for resnet, attn in zip(self.resnets, self.attentions):
1359
1360
1361
            hidden_states = resnet(hidden_states, temb=None, scale=scale)
            cross_attention_kwargs = {"scale": scale}
            hidden_states = attn(hidden_states, **cross_attention_kwargs)
1362
1363
1364

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1365
                hidden_states = downsampler(hidden_states, scale)
1366
1367
1368
1369

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
1370
class AttnSkipDownBlock2D(nn.Module):
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
    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,
1382
        attention_head_dim=1,
1383
1384
1385
1386
1387
1388
1389
        output_scale_factor=np.sqrt(2.0),
        add_downsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        self.resnets = nn.ModuleList([])

1390
1391
1392
1393
1394
1395
        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

1396
1397
1398
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
1399
                ResnetBlock2D(
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
                    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(
1414
                Attention(
1415
                    out_channels,
1416
1417
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1418
1419
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
1420
1421
1422
1423
1424
                    norm_num_groups=32,
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1425
1426
1427
1428
                )
            )

        if add_downsample:
1429
            self.resnet_down = ResnetBlock2D(
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
                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,
1440
                use_in_shortcut=True,
1441
1442
1443
                down=True,
                kernel="fir",
            )
1444
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1445
1446
1447
1448
1449
1450
            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

1451
    def forward(self, hidden_states, temb=None, skip_sample=None, scale: float = 1.0):
1452
1453
1454
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
1455
1456
1457
            hidden_states = resnet(hidden_states, temb, scale=scale)
            cross_attention_kwargs = {"scale": scale}
            hidden_states = attn(hidden_states, **cross_attention_kwargs)
1458
1459
1460
            output_states += (hidden_states,)

        if self.downsamplers is not None:
1461
            hidden_states = self.resnet_down(hidden_states, temb, scale=scale)
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
            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
1472
class SkipDownBlock2D(nn.Module):
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
    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(
1494
                ResnetBlock2D(
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
                    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:
1510
            self.resnet_down = ResnetBlock2D(
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
                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,
1521
                use_in_shortcut=True,
1522
1523
1524
                down=True,
                kernel="fir",
            )
1525
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1526
1527
1528
1529
1530
1531
            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

1532
    def forward(self, hidden_states, temb=None, skip_sample=None, scale: float = 1.0):
1533
1534
1535
        output_states = ()

        for resnet in self.resnets:
1536
            hidden_states = resnet(hidden_states, temb, scale)
1537
1538
1539
            output_states += (hidden_states,)

        if self.downsamplers is not None:
1540
            hidden_states = self.resnet_down(hidden_states, temb, scale)
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
            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
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
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,
1566
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
    ):
        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,
1585
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
                )
            )

        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,
1605
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1606
1607
1608
1609
1610
1611
1612
1613
1614
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

1615
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
Will Berman's avatar
Will Berman committed
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
        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

1627
1628
1629
1630
1631
1632
1633
1634
                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
1635
            else:
1636
                hidden_states = resnet(hidden_states, temb, scale)
Will Berman's avatar
Will Berman committed
1637

1638
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1639
1640
1641

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1642
                hidden_states = downsampler(hidden_states, temb, scale)
Will Berman's avatar
Will Berman committed
1643

1644
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661

        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,
1662
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
1663
1664
1665
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_downsample=True,
1666
        skip_time_act=False,
1667
        only_cross_attention=False,
1668
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
1669
1670
1671
1672
1673
1674
1675
1676
    ):
        super().__init__()

        self.has_cross_attention = True

        resnets = []
        attentions = []

1677
1678
        self.attention_head_dim = attention_head_dim
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693

        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,
1694
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1695
1696
                )
            )
1697
1698
1699
1700
1701

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

Will Berman's avatar
Will Berman committed
1702
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
1703
                Attention(
Will Berman's avatar
Will Berman committed
1704
1705
1706
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
1707
                    dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
1708
1709
1710
1711
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
1712
                    only_cross_attention=only_cross_attention,
1713
                    cross_attention_norm=cross_attention_norm,
1714
                    processor=processor,
Will Berman's avatar
Will Berman committed
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
                )
            )
        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,
1734
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1735
1736
1737
1738
1739
1740
1741
1742
1743
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

1744
    def forward(
1745
1746
1747
1748
1749
1750
1751
        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,
1752
    ):
Will Berman's avatar
Will Berman committed
1753
        output_states = ()
1754
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
Will Berman's avatar
Will Berman committed
1755

1756
1757
        lora_scale = cross_attention_kwargs.get("scale", 1.0)

1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
        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
1769
        for resnet, attn in zip(self.resnets, self.attentions):
1770
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
1771

1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
                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
1782
                hidden_states = attn(
1783
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
1784
1785
1786
1787
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=mask,
                    **cross_attention_kwargs,
                )
1788
            else:
1789
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
1790
1791
1792
1793

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
1794
                    attention_mask=mask,
1795
1796
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
1797

1798
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1799
1800
1801

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
1802
                hidden_states = downsampler(hidden_states, temb, scale=lora_scale)
Will Berman's avatar
Will Berman committed
1803

1804
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1805
1806
1807
1808

        return hidden_states, output_states


1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
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

1855
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
        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

1867
1868
1869
1870
1871
1872
1873
1874
                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
                    )
1875
            else:
1876
                hidden_states = resnet(hidden_states, temb, scale)
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897

            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,
1898
        attention_head_dim: int = 64,
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
        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,
1931
1932
                    out_channels // attention_head_dim,
                    attention_head_dim,
1933
1934
1935
1936
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
1937
                    cross_attention_norm="layer_norm",
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
                    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(
1953
1954
1955
1956
1957
1958
1959
        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,
1960
1961
    ):
        output_states = ()
1962
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975

        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

1976
1977
1978
1979
1980
1981
1982
                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
1983
                hidden_states = attn(
1984
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
1985
1986
1987
1988
1989
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
1990
                )
1991
            else:
1992
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
1993
1994
1995
1996
1997
1998
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
1999
                    encoder_attention_mask=encoder_attention_mask,
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
                )

            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
2014
class AttnUpBlock2D(nn.Module):
2015
2016
2017
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
2018
2019
        prev_output_channel: int,
        out_channels: int,
2020
        temb_channels: int,
2021
        resolution_idx: int = None,
2022
2023
2024
2025
2026
2027
2028
        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,
2029
        attention_head_dim=1,
2030
        output_scale_factor=1.0,
2031
        upsample_type="conv",
2032
2033
2034
2035
2036
    ):
        super().__init__()
        resnets = []
        attentions = []

2037
2038
        self.upsample_type = upsample_type

2039
2040
2041
2042
2043
2044
        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

2045
        for i in range(num_layers):
Patrick von Platen's avatar
Patrick von Platen committed
2046
2047
2048
            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

2049
            resnets.append(
2050
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2051
2052
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
                    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(
2064
                Attention(
Patrick von Platen's avatar
Patrick von Platen committed
2065
                    out_channels,
2066
2067
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
2068
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
2069
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
2070
                    norm_num_groups=resnet_groups,
2071
2072
2073
2074
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
2075
2076
2077
2078
2079
2080
                )
            )

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

2081
        if upsample_type == "conv":
Patrick von Platen's avatar
Patrick von Platen committed
2082
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
        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,
                    )
                ]
            )
2101
2102
2103
        else:
            self.upsamplers = None

2104
2105
        self.resolution_idx = resolution_idx

2106
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, scale: float = 1.0):
2107
2108
2109
2110
2111
2112
        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)

2113
2114
2115
            hidden_states = resnet(hidden_states, temb, scale=scale)
            cross_attention_kwargs = {"scale": scale}
            hidden_states = attn(hidden_states, **cross_attention_kwargs)
2116
2117
2118

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2119
                if self.upsample_type == "resnet":
2120
                    hidden_states = upsampler(hidden_states, temb=temb, scale=scale)
2121
                else:
2122
                    hidden_states = upsampler(hidden_states, scale=scale)
2123
2124
2125
2126

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2127
class CrossAttnUpBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
2128
2129
2130
2131
2132
2133
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
2134
        resolution_idx: int = None,
Patrick von Platen's avatar
Patrick von Platen committed
2135
2136
        dropout: float = 0.0,
        num_layers: int = 1,
2137
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
2138
2139
2140
2141
2142
        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,
2143
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
2144
2145
2146
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2147
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
2148
        use_linear_projection=False,
2149
        only_cross_attention=False,
2150
        upcast_attention=False,
2151
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
2152
2153
2154
2155
2156
    ):
        super().__init__()
        resnets = []
        attentions = []

2157
        self.has_cross_attention = True
2158
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
2159
2160
2161
2162
2163
2164

        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(
2165
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
                    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,
                )
            )
2178
2179
2180
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
2181
2182
                        num_attention_heads,
                        out_channels // num_attention_heads,
2183
                        in_channels=out_channels,
2184
                        num_layers=transformer_layers_per_block,
2185
2186
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
2187
                        use_linear_projection=use_linear_projection,
2188
                        only_cross_attention=only_cross_attention,
2189
                        upcast_attention=upcast_attention,
2190
                        attention_type=attention_type,
2191
2192
2193
2194
2195
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
2196
2197
                        num_attention_heads,
                        out_channels // num_attention_heads,
2198
2199
2200
2201
2202
                        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
2203
2204
2205
2206
2207
2208
2209
2210
2211
                )
        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

2212
        self.gradient_checkpointing = False
2213
        self.resolution_idx = resolution_idx
2214
2215
2216

    def forward(
        self,
2217
2218
2219
2220
2221
2222
2223
2224
        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,
2225
    ):
2226
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
2227
2228
2229
2230
2231
2232
        is_freeu_enabled = (
            getattr(self, "s1", None)
            and getattr(self, "s2", None)
            and getattr(self, "b1", None)
            and getattr(self, "b2", None)
        )
2233

Patrick von Platen's avatar
Patrick von Platen committed
2234
2235
2236
2237
        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]
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250

            # 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
2251
2252
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

2253
2254
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
2255
                def create_custom_forward(module, return_dict=None):
2256
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
2257
2258
2259
2260
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
2261
2262
2263

                    return custom_forward

2264
2265
2266
2267
2268
2269
2270
                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
2271
                hidden_states = attn(
2272
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
2273
2274
2275
2276
2277
                    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,
2278
                )[0]
2279
            else:
2280
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
2281
2282
2283
2284
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
2285
2286
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
2287
2288
                    return_dict=False,
                )[0]
Patrick von Platen's avatar
Patrick von Platen committed
2289
2290
2291

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2292
                hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
Patrick von Platen's avatar
Patrick von Platen committed
2293
2294
2295
2296

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2297
class UpBlock2D(nn.Module):
2298
2299
2300
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
2301
2302
        prev_output_channel: int,
        out_channels: int,
2303
        temb_channels: int,
2304
        resolution_idx: int = None,
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
        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
2319
2320
2321
            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

2322
            resnets.append(
2323
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2324
2325
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
                    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
2340
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
2341
2342
2343
        else:
            self.upsamplers = None

2344
        self.gradient_checkpointing = False
2345
        self.resolution_idx = resolution_idx
2346

2347
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, scale: float = 1.0):
2348
2349
2350
2351
2352
2353
2354
        is_freeu_enabled = (
            getattr(self, "s1", None)
            and getattr(self, "s2", None)
            and getattr(self, "b1", None)
            and getattr(self, "b2", None)
        )

2355
2356
2357
2358
        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]
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371

            # 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,
                )

2372
2373
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

2374
2375
2376
2377
2378
2379
2380
2381
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

2382
2383
2384
2385
2386
2387
2388
2389
                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
                    )
2390
            else:
2391
                hidden_states = resnet(hidden_states, temb, scale=scale)
2392
2393
2394

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2395
                hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
2396
2397

        return hidden_states
2398
2399


2400
2401
2402
2403
2404
class UpDecoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
2405
        resolution_idx: int = None,
2406
2407
2408
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
YiYi Xu's avatar
YiYi Xu committed
2409
        resnet_time_scale_shift: str = "default",  # default, spatial
2410
2411
2412
2413
2414
        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
2415
        temb_channels=None,
2416
2417
2418
2419
2420
2421
2422
2423
    ):
        super().__init__()
        resnets = []

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

            resnets.append(
2424
                ResnetBlock2D(
2425
2426
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2427
                    temb_channels=temb_channels,
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
                    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

2445
2446
        self.resolution_idx = resolution_idx

2447
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
2448
        for resnet in self.resnets:
2449
            hidden_states = resnet(hidden_states, temb=temb, scale=scale)
2450
2451
2452
2453
2454
2455
2456
2457

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

        return hidden_states


2458
2459
2460
2461
2462
class AttnUpDecoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
2463
        resolution_idx: int = None,
2464
2465
2466
2467
2468
2469
2470
        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,
2471
        attention_head_dim=1,
2472
2473
        output_scale_factor=1.0,
        add_upsample=True,
YiYi Xu's avatar
YiYi Xu committed
2474
        temb_channels=None,
2475
2476
2477
2478
2479
    ):
        super().__init__()
        resnets = []
        attentions = []

2480
2481
2482
2483
2484
2485
        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

2486
2487
2488
2489
        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
2490
                ResnetBlock2D(
2491
2492
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2493
                    temb_channels=temb_channels,
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
                    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(
2504
                Attention(
2505
                    out_channels,
2506
2507
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
2508
2509
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
2510
                    norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
YiYi Xu's avatar
YiYi Xu committed
2511
                    spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
2512
2513
2514
2515
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
                )
            )

        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

2527
2528
        self.resolution_idx = resolution_idx

2529
    def forward(self, hidden_states, temb=None, scale: float = 1.0):
2530
        for resnet, attn in zip(self.resnets, self.attentions):
2531
2532
2533
            hidden_states = resnet(hidden_states, temb=temb, scale=scale)
            cross_attention_kwargs = {"scale": scale}
            hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs)
2534
2535
2536

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2537
                hidden_states = upsampler(hidden_states, scale=scale)
2538
2539
2540
2541

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2542
class AttnSkipUpBlock2D(nn.Module):
2543
2544
2545
2546
2547
2548
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
2549
        resolution_idx: int = None,
2550
2551
2552
2553
2554
2555
        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,
2556
        attention_head_dim=1,
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
        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(
2569
                ResnetBlock2D(
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
                    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,
                )
            )

2584
2585
2586
2587
2588
2589
        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

2590
        self.attentions.append(
2591
            Attention(
2592
                out_channels,
2593
2594
                heads=out_channels // attention_head_dim,
                dim_head=attention_head_dim,
2595
2596
                rescale_output_factor=output_scale_factor,
                eps=resnet_eps,
2597
2598
2599
2600
2601
                norm_num_groups=32,
                residual_connection=True,
                bias=True,
                upcast_softmax=True,
                _from_deprecated_attn_block=True,
2602
2603
2604
2605
2606
            )
        )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
2607
            self.resnet_up = ResnetBlock2D(
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
                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,
2619
                use_in_shortcut=True,
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
                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

2634
2635
        self.resolution_idx = resolution_idx

2636
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None, scale: float = 1.0):
2637
2638
2639
2640
2641
2642
        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)

2643
            hidden_states = resnet(hidden_states, temb, scale=scale)
2644

2645
2646
        cross_attention_kwargs = {"scale": scale}
        hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs)
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659

        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

2660
            hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
2661
2662
2663
2664

        return hidden_states, skip_sample


Patrick von Platen's avatar
Patrick von Platen committed
2665
class SkipUpBlock2D(nn.Module):
2666
2667
2668
2669
2670
2671
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
2672
        resolution_idx: int = None,
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
        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(
2691
                ResnetBlock2D(
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
                    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:
2708
            self.resnet_up = ResnetBlock2D(
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
                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,
2720
                use_in_shortcut=True,
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
                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

2735
2736
        self.resolution_idx = resolution_idx

2737
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None, scale: float = 1.0):
2738
2739
2740
2741
2742
2743
        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)

2744
            hidden_states = resnet(hidden_states, temb, scale=scale)
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757

        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

2758
            hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
2759
2760

        return hidden_states, skip_sample
Will Berman's avatar
Will Berman committed
2761
2762
2763
2764
2765
2766
2767
2768
2769


class ResnetUpsampleBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
2770
        resolution_idx: int = None,
Will Berman's avatar
Will Berman committed
2771
2772
2773
2774
2775
2776
2777
2778
2779
        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,
2780
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
    ):
        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,
2801
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
                )
            )

        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,
2821
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2822
2823
2824
2825
2826
2827
2828
2829
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False
2830
        self.resolution_idx = resolution_idx
Will Berman's avatar
Will Berman committed
2831

2832
    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
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
        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

2847
2848
2849
2850
2851
2852
2853
2854
                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
2855
            else:
2856
                hidden_states = resnet(hidden_states, temb, scale=scale)
Will Berman's avatar
Will Berman committed
2857
2858
2859

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2860
                hidden_states = upsampler(hidden_states, temb, scale=scale)
Will Berman's avatar
Will Berman committed
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871

        return hidden_states


class SimpleCrossAttnUpBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
2872
        resolution_idx: int = None,
Will Berman's avatar
Will Berman committed
2873
2874
2875
2876
2877
2878
2879
        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,
2880
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
2881
2882
2883
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2884
        skip_time_act=False,
2885
        only_cross_attention=False,
2886
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
2887
2888
2889
2890
2891
2892
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
2893
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
2894

2895
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912

        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,
2913
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2914
2915
                )
            )
2916
2917
2918
2919
2920

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

Will Berman's avatar
Will Berman committed
2921
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
2922
                Attention(
Will Berman's avatar
Will Berman committed
2923
2924
2925
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
2926
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
2927
2928
2929
2930
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
2931
                    only_cross_attention=only_cross_attention,
2932
                    cross_attention_norm=cross_attention_norm,
2933
                    processor=processor,
Will Berman's avatar
Will Berman committed
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
                )
            )
        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,
2953
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2954
2955
2956
2957
2958
2959
2960
2961
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False
2962
        self.resolution_idx = resolution_idx
Will Berman's avatar
Will Berman committed
2963
2964
2965

    def forward(
        self,
2966
2967
2968
2969
2970
2971
2972
2973
        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
2974
    ):
2975
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
2976

2977
        lora_scale = cross_attention_kwargs.get("scale", 1.0)
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
        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
2989
2990
2991
2992
2993
2994
2995
        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)

2996
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
2997

2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
                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
3008
                hidden_states = attn(
3009
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
3010
3011
3012
3013
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=mask,
                    **cross_attention_kwargs,
                )
3014
            else:
3015
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
3016
3017
3018
3019

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
3020
                    attention_mask=mask,
3021
3022
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
3023
3024
3025

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
3026
                hidden_states = upsampler(hidden_states, temb, scale=lora_scale)
Will Berman's avatar
Will Berman committed
3027
3028

        return hidden_states
3029
3030
3031
3032
3033
3034
3035
3036


class KUpBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
3037
        resolution_idx: int,
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
        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
3079
        self.resolution_idx = resolution_idx
3080

3081
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, scale: float = 1.0):
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
        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

3095
3096
3097
3098
3099
3100
3101
3102
                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
                    )
3103
            else:
3104
                hidden_states = resnet(hidden_states, temb, scale=scale)
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118

        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,
3119
        resolution_idx: int,
3120
3121
3122
3123
3124
        dropout: float = 0.0,
        num_layers: int = 4,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
        resnet_group_size: int = 32,
3125
        attention_head_dim=1,  # attention dim_head
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
        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
3139
        self.attention_head_dim = attention_head_dim
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174

        # 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,
3175
                    k_out_channels // attention_head_dim
3176
                    if (i == num_layers - 1)
3177
3178
                    else out_channels // attention_head_dim,
                    attention_head_dim,
3179
3180
3181
3182
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
3183
                    cross_attention_norm="layer_norm",
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
                    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
3197
        self.resolution_idx = resolution_idx
3198
3199
3200

    def forward(
        self,
3201
3202
3203
3204
3205
3206
3207
3208
        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,
3209
3210
3211
3212
3213
    ):
        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)

3214
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
        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

3227
3228
3229
3230
3231
3232
3233
                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
3234
                hidden_states = attn(
3235
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
3236
3237
3238
3239
3240
3241
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
                )
3242
            else:
3243
                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
3244
3245
3246
3247
3248
3249
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
3250
                    encoder_attention_mask=encoder_attention_mask,
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
                )

        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,
3289
        cross_attention_norm: Optional[str] = None,
3290
3291
3292
3293
3294
3295
3296
3297
        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
3298
            self.attn1 = Attention(
3299
3300
3301
3302
3303
3304
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                cross_attention_dim=None,
3305
                cross_attention_norm=None,
3306
3307
3308
3309
            )

        # 2. Cross-Attn
        self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
Patrick von Platen's avatar
Patrick von Platen committed
3310
        self.attn2 = Attention(
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
            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,
3329
3330
3331
3332
3333
3334
3335
3336
        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,
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
    ):
        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,
3350
                attention_mask=attention_mask,
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
                **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,
3365
            attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
3366
3367
3368
3369
3370
3371
3372
            **cross_attention_kwargs,
        )
        attn_output = self._to_4d(attn_output, height, weight)

        hidden_states = attn_output + hidden_states

        return hidden_states