File size: 24,596 Bytes
ffead1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
# coding=utf-8
# Copyright 2023 HuggingFace Inc.
#
# 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.


import unittest

import numpy as np
import torch
from torch import nn

from diffusers.models.attention import GEGLU, AdaLayerNorm, ApproximateGELU, AttentionBlock
from diffusers.models.embeddings import get_timestep_embedding
from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
from diffusers.models.transformer_2d import Transformer2DModel
from diffusers.utils import torch_device


torch.backends.cuda.matmul.allow_tf32 = False


class EmbeddingsTests(unittest.TestCase):
    def test_timestep_embeddings(self):
        embedding_dim = 256
        timesteps = torch.arange(16)

        t1 = get_timestep_embedding(timesteps, embedding_dim)

        # first vector should always be composed only of 0's and 1's
        assert (t1[0, : embedding_dim // 2] - 0).abs().sum() < 1e-5
        assert (t1[0, embedding_dim // 2 :] - 1).abs().sum() < 1e-5

        # last element of each vector should be one
        assert (t1[:, -1] - 1).abs().sum() < 1e-5

        # For large embeddings (e.g. 128) the frequency of every vector is higher
        # than the previous one which means that the gradients of later vectors are
        # ALWAYS higher than the previous ones
        grad_mean = np.abs(np.gradient(t1, axis=-1)).mean(axis=1)

        prev_grad = 0.0
        for grad in grad_mean:
            assert grad > prev_grad
            prev_grad = grad

    def test_timestep_defaults(self):
        embedding_dim = 16
        timesteps = torch.arange(10)

        t1 = get_timestep_embedding(timesteps, embedding_dim)
        t2 = get_timestep_embedding(
            timesteps, embedding_dim, flip_sin_to_cos=False, downscale_freq_shift=1, max_period=10_000
        )

        assert torch.allclose(t1.cpu(), t2.cpu(), 1e-3)

    def test_timestep_flip_sin_cos(self):
        embedding_dim = 16
        timesteps = torch.arange(10)

        t1 = get_timestep_embedding(timesteps, embedding_dim, flip_sin_to_cos=True)
        t1 = torch.cat([t1[:, embedding_dim // 2 :], t1[:, : embedding_dim // 2]], dim=-1)

        t2 = get_timestep_embedding(timesteps, embedding_dim, flip_sin_to_cos=False)

        assert torch.allclose(t1.cpu(), t2.cpu(), 1e-3)

    def test_timestep_downscale_freq_shift(self):
        embedding_dim = 16
        timesteps = torch.arange(10)

        t1 = get_timestep_embedding(timesteps, embedding_dim, downscale_freq_shift=0)
        t2 = get_timestep_embedding(timesteps, embedding_dim, downscale_freq_shift=1)

        # get cosine half (vectors that are wrapped into cosine)
        cosine_half = (t1 - t2)[:, embedding_dim // 2 :]

        # cosine needs to be negative
        assert (np.abs((cosine_half <= 0).numpy()) - 1).sum() < 1e-5

    def test_sinoid_embeddings_hardcoded(self):
        embedding_dim = 64
        timesteps = torch.arange(128)

        # standard unet, score_vde
        t1 = get_timestep_embedding(timesteps, embedding_dim, downscale_freq_shift=1, flip_sin_to_cos=False)
        # glide, ldm
        t2 = get_timestep_embedding(timesteps, embedding_dim, downscale_freq_shift=0, flip_sin_to_cos=True)
        # grad-tts
        t3 = get_timestep_embedding(timesteps, embedding_dim, scale=1000)

        assert torch.allclose(
            t1[23:26, 47:50].flatten().cpu(),
            torch.tensor([0.9646, 0.9804, 0.9892, 0.9615, 0.9787, 0.9882, 0.9582, 0.9769, 0.9872]),
            1e-3,
        )
        assert torch.allclose(
            t2[23:26, 47:50].flatten().cpu(),
            torch.tensor([0.3019, 0.2280, 0.1716, 0.3146, 0.2377, 0.1790, 0.3272, 0.2474, 0.1864]),
            1e-3,
        )
        assert torch.allclose(
            t3[23:26, 47:50].flatten().cpu(),
            torch.tensor([-0.9801, -0.9464, -0.9349, -0.3952, 0.8887, -0.9709, 0.5299, -0.2853, -0.9927]),
            1e-3,
        )


class Upsample2DBlockTests(unittest.TestCase):
    def test_upsample_default(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 32, 32)
        upsample = Upsample2D(channels=32, use_conv=False)
        with torch.no_grad():
            upsampled = upsample(sample)

        assert upsampled.shape == (1, 32, 64, 64)
        output_slice = upsampled[0, -1, -3:, -3:]
        expected_slice = torch.tensor([-0.2173, -1.2079, -1.2079, 0.2952, 1.1254, 1.1254, 0.2952, 1.1254, 1.1254])
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_upsample_with_conv(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 32, 32)
        upsample = Upsample2D(channels=32, use_conv=True)
        with torch.no_grad():
            upsampled = upsample(sample)

        assert upsampled.shape == (1, 32, 64, 64)
        output_slice = upsampled[0, -1, -3:, -3:]
        expected_slice = torch.tensor([0.7145, 1.3773, 0.3492, 0.8448, 1.0839, -0.3341, 0.5956, 0.1250, -0.4841])
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_upsample_with_conv_out_dim(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 32, 32)
        upsample = Upsample2D(channels=32, use_conv=True, out_channels=64)
        with torch.no_grad():
            upsampled = upsample(sample)

        assert upsampled.shape == (1, 64, 64, 64)
        output_slice = upsampled[0, -1, -3:, -3:]
        expected_slice = torch.tensor([0.2703, 0.1656, -0.2538, -0.0553, -0.2984, 0.1044, 0.1155, 0.2579, 0.7755])
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_upsample_with_transpose(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 32, 32)
        upsample = Upsample2D(channels=32, use_conv=False, use_conv_transpose=True)
        with torch.no_grad():
            upsampled = upsample(sample)

        assert upsampled.shape == (1, 32, 64, 64)
        output_slice = upsampled[0, -1, -3:, -3:]
        expected_slice = torch.tensor([-0.3028, -0.1582, 0.0071, 0.0350, -0.4799, -0.1139, 0.1056, -0.1153, -0.1046])
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)


class Downsample2DBlockTests(unittest.TestCase):
    def test_downsample_default(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64)
        downsample = Downsample2D(channels=32, use_conv=False)
        with torch.no_grad():
            downsampled = downsample(sample)

        assert downsampled.shape == (1, 32, 32, 32)
        output_slice = downsampled[0, -1, -3:, -3:]
        expected_slice = torch.tensor([-0.0513, -0.3889, 0.0640, 0.0836, -0.5460, -0.0341, -0.0169, -0.6967, 0.1179])
        max_diff = (output_slice.flatten() - expected_slice).abs().sum().item()
        assert max_diff <= 1e-3
        # assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-1)

    def test_downsample_with_conv(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64)
        downsample = Downsample2D(channels=32, use_conv=True)
        with torch.no_grad():
            downsampled = downsample(sample)

        assert downsampled.shape == (1, 32, 32, 32)
        output_slice = downsampled[0, -1, -3:, -3:]

        expected_slice = torch.tensor(
            [0.9267, 0.5878, 0.3337, 1.2321, -0.1191, -0.3984, -0.7532, -0.0715, -0.3913],
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_downsample_with_conv_pad1(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64)
        downsample = Downsample2D(channels=32, use_conv=True, padding=1)
        with torch.no_grad():
            downsampled = downsample(sample)

        assert downsampled.shape == (1, 32, 32, 32)
        output_slice = downsampled[0, -1, -3:, -3:]
        expected_slice = torch.tensor([0.9267, 0.5878, 0.3337, 1.2321, -0.1191, -0.3984, -0.7532, -0.0715, -0.3913])
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_downsample_with_conv_out_dim(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64)
        downsample = Downsample2D(channels=32, use_conv=True, out_channels=16)
        with torch.no_grad():
            downsampled = downsample(sample)

        assert downsampled.shape == (1, 16, 32, 32)
        output_slice = downsampled[0, -1, -3:, -3:]
        expected_slice = torch.tensor([-0.6586, 0.5985, 0.0721, 0.1256, -0.1492, 0.4436, -0.2544, 0.5021, 1.1522])
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)


class ResnetBlock2DTests(unittest.TestCase):
    def test_resnet_default(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        temb = torch.randn(1, 128).to(torch_device)
        resnet_block = ResnetBlock2D(in_channels=32, temb_channels=128).to(torch_device)
        with torch.no_grad():
            output_tensor = resnet_block(sample, temb)

        assert output_tensor.shape == (1, 32, 64, 64)
        output_slice = output_tensor[0, -1, -3:, -3:]
        expected_slice = torch.tensor(
            [-1.9010, -0.2974, -0.8245, -1.3533, 0.8742, -0.9645, -2.0584, 1.3387, -0.4746], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_restnet_with_use_in_shortcut(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        temb = torch.randn(1, 128).to(torch_device)
        resnet_block = ResnetBlock2D(in_channels=32, temb_channels=128, use_in_shortcut=True).to(torch_device)
        with torch.no_grad():
            output_tensor = resnet_block(sample, temb)

        assert output_tensor.shape == (1, 32, 64, 64)
        output_slice = output_tensor[0, -1, -3:, -3:]
        expected_slice = torch.tensor(
            [0.2226, -1.0791, -0.1629, 0.3659, -0.2889, -1.2376, 0.0582, 0.9206, 0.0044], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_resnet_up(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        temb = torch.randn(1, 128).to(torch_device)
        resnet_block = ResnetBlock2D(in_channels=32, temb_channels=128, up=True).to(torch_device)
        with torch.no_grad():
            output_tensor = resnet_block(sample, temb)

        assert output_tensor.shape == (1, 32, 128, 128)
        output_slice = output_tensor[0, -1, -3:, -3:]
        expected_slice = torch.tensor(
            [1.2130, -0.8753, -0.9027, 1.5783, -0.5362, -0.5001, 1.0726, -0.7732, -0.4182], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_resnet_down(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        temb = torch.randn(1, 128).to(torch_device)
        resnet_block = ResnetBlock2D(in_channels=32, temb_channels=128, down=True).to(torch_device)
        with torch.no_grad():
            output_tensor = resnet_block(sample, temb)

        assert output_tensor.shape == (1, 32, 32, 32)
        output_slice = output_tensor[0, -1, -3:, -3:]
        expected_slice = torch.tensor(
            [-0.3002, -0.7135, 0.1359, 0.0561, -0.7935, 0.0113, -0.1766, -0.6714, -0.0436], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_restnet_with_kernel_fir(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        temb = torch.randn(1, 128).to(torch_device)
        resnet_block = ResnetBlock2D(in_channels=32, temb_channels=128, kernel="fir", down=True).to(torch_device)
        with torch.no_grad():
            output_tensor = resnet_block(sample, temb)

        assert output_tensor.shape == (1, 32, 32, 32)
        output_slice = output_tensor[0, -1, -3:, -3:]
        expected_slice = torch.tensor(
            [-0.0934, -0.5729, 0.0909, -0.2710, -0.5044, 0.0243, -0.0665, -0.5267, -0.3136], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_restnet_with_kernel_sde_vp(self):
        torch.manual_seed(0)
        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        temb = torch.randn(1, 128).to(torch_device)
        resnet_block = ResnetBlock2D(in_channels=32, temb_channels=128, kernel="sde_vp", down=True).to(torch_device)
        with torch.no_grad():
            output_tensor = resnet_block(sample, temb)

        assert output_tensor.shape == (1, 32, 32, 32)
        output_slice = output_tensor[0, -1, -3:, -3:]
        expected_slice = torch.tensor(
            [-0.3002, -0.7135, 0.1359, 0.0561, -0.7935, 0.0113, -0.1766, -0.6714, -0.0436], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)


class AttentionBlockTests(unittest.TestCase):
    @unittest.skipIf(
        torch_device == "mps", "Matmul crashes on MPS, see https://github.com/pytorch/pytorch/issues/84039"
    )
    def test_attention_block_default(self):
        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        attentionBlock = AttentionBlock(
            channels=32,
            num_head_channels=1,
            rescale_output_factor=1.0,
            eps=1e-6,
            norm_num_groups=32,
        ).to(torch_device)
        with torch.no_grad():
            attention_scores = attentionBlock(sample)

        assert attention_scores.shape == (1, 32, 64, 64)
        output_slice = attention_scores[0, -1, -3:, -3:]

        expected_slice = torch.tensor(
            [-1.4975, -0.0038, -0.7847, -1.4567, 1.1220, -0.8962, -1.7394, 1.1319, -0.5427], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_attention_block_sd(self):
        # This version uses SD params and is compatible with mps
        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        sample = torch.randn(1, 512, 64, 64).to(torch_device)
        attentionBlock = AttentionBlock(
            channels=512,
            rescale_output_factor=1.0,
            eps=1e-6,
            norm_num_groups=32,
        ).to(torch_device)
        with torch.no_grad():
            attention_scores = attentionBlock(sample)

        assert attention_scores.shape == (1, 512, 64, 64)
        output_slice = attention_scores[0, -1, -3:, -3:]

        expected_slice = torch.tensor(
            [-0.6621, -0.0156, -3.2766, 0.8025, -0.8609, 0.2820, 0.0905, -1.1179, -3.2126], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)


class Transformer2DModelTests(unittest.TestCase):
    def test_spatial_transformer_default(self):
        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        spatial_transformer_block = Transformer2DModel(
            in_channels=32,
            num_attention_heads=1,
            attention_head_dim=32,
            dropout=0.0,
            cross_attention_dim=None,
        ).to(torch_device)
        with torch.no_grad():
            attention_scores = spatial_transformer_block(sample).sample

        assert attention_scores.shape == (1, 32, 64, 64)
        output_slice = attention_scores[0, -1, -3:, -3:]

        expected_slice = torch.tensor(
            [-1.9455, -0.0066, -1.3933, -1.5878, 0.5325, -0.6486, -1.8648, 0.7515, -0.9689], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_spatial_transformer_cross_attention_dim(self):
        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        sample = torch.randn(1, 64, 64, 64).to(torch_device)
        spatial_transformer_block = Transformer2DModel(
            in_channels=64,
            num_attention_heads=2,
            attention_head_dim=32,
            dropout=0.0,
            cross_attention_dim=64,
        ).to(torch_device)
        with torch.no_grad():
            context = torch.randn(1, 4, 64).to(torch_device)
            attention_scores = spatial_transformer_block(sample, context).sample

        assert attention_scores.shape == (1, 64, 64, 64)
        output_slice = attention_scores[0, -1, -3:, -3:]

        expected_slice = torch.tensor(
            [-0.2555, -0.8877, -2.4739, -2.2251, 1.2714, 0.0807, -0.4161, -1.6408, -0.0471], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_spatial_transformer_timestep(self):
        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        num_embeds_ada_norm = 5

        sample = torch.randn(1, 64, 64, 64).to(torch_device)
        spatial_transformer_block = Transformer2DModel(
            in_channels=64,
            num_attention_heads=2,
            attention_head_dim=32,
            dropout=0.0,
            cross_attention_dim=64,
            num_embeds_ada_norm=num_embeds_ada_norm,
        ).to(torch_device)
        with torch.no_grad():
            timestep_1 = torch.tensor(1, dtype=torch.long).to(torch_device)
            timestep_2 = torch.tensor(2, dtype=torch.long).to(torch_device)
            attention_scores_1 = spatial_transformer_block(sample, timestep=timestep_1).sample
            attention_scores_2 = spatial_transformer_block(sample, timestep=timestep_2).sample

        assert attention_scores_1.shape == (1, 64, 64, 64)
        assert attention_scores_2.shape == (1, 64, 64, 64)

        output_slice_1 = attention_scores_1[0, -1, -3:, -3:]
        output_slice_2 = attention_scores_2[0, -1, -3:, -3:]

        expected_slice_1 = torch.tensor(
            [-0.1874, -0.9704, -1.4290, -1.3357, 1.5138, 0.3036, -0.0976, -1.1667, 0.1283], device=torch_device
        )
        expected_slice_2 = torch.tensor(
            [-0.3493, -1.0924, -1.6161, -1.5016, 1.4245, 0.1367, -0.2526, -1.3109, -0.0547], device=torch_device
        )

        assert torch.allclose(output_slice_1.flatten(), expected_slice_1, atol=1e-3)
        assert torch.allclose(output_slice_2.flatten(), expected_slice_2, atol=1e-3)

    def test_spatial_transformer_dropout(self):
        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        sample = torch.randn(1, 32, 64, 64).to(torch_device)
        spatial_transformer_block = (
            Transformer2DModel(
                in_channels=32,
                num_attention_heads=2,
                attention_head_dim=16,
                dropout=0.3,
                cross_attention_dim=None,
            )
            .to(torch_device)
            .eval()
        )
        with torch.no_grad():
            attention_scores = spatial_transformer_block(sample).sample

        assert attention_scores.shape == (1, 32, 64, 64)
        output_slice = attention_scores[0, -1, -3:, -3:]

        expected_slice = torch.tensor(
            [-1.9380, -0.0083, -1.3771, -1.5819, 0.5209, -0.6441, -1.8545, 0.7563, -0.9615], device=torch_device
        )
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    @unittest.skipIf(torch_device == "mps", "MPS does not support float64")
    def test_spatial_transformer_discrete(self):
        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        num_embed = 5

        sample = torch.randint(0, num_embed, (1, 32)).to(torch_device)
        spatial_transformer_block = (
            Transformer2DModel(
                num_attention_heads=1,
                attention_head_dim=32,
                num_vector_embeds=num_embed,
                sample_size=16,
            )
            .to(torch_device)
            .eval()
        )

        with torch.no_grad():
            attention_scores = spatial_transformer_block(sample).sample

        assert attention_scores.shape == (1, num_embed - 1, 32)

        output_slice = attention_scores[0, -2:, -3:]

        expected_slice = torch.tensor([-1.7648, -1.0241, -2.0985, -1.8035, -1.6404, -1.2098], device=torch_device)
        assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)

    def test_spatial_transformer_default_norm_layers(self):
        spatial_transformer_block = Transformer2DModel(num_attention_heads=1, attention_head_dim=32, in_channels=32)

        assert spatial_transformer_block.transformer_blocks[0].norm1.__class__ == nn.LayerNorm
        assert spatial_transformer_block.transformer_blocks[0].norm3.__class__ == nn.LayerNorm

    def test_spatial_transformer_ada_norm_layers(self):
        spatial_transformer_block = Transformer2DModel(
            num_attention_heads=1,
            attention_head_dim=32,
            in_channels=32,
            num_embeds_ada_norm=5,
        )

        assert spatial_transformer_block.transformer_blocks[0].norm1.__class__ == AdaLayerNorm
        assert spatial_transformer_block.transformer_blocks[0].norm3.__class__ == nn.LayerNorm

    def test_spatial_transformer_default_ff_layers(self):
        spatial_transformer_block = Transformer2DModel(
            num_attention_heads=1,
            attention_head_dim=32,
            in_channels=32,
        )

        assert spatial_transformer_block.transformer_blocks[0].ff.net[0].__class__ == GEGLU
        assert spatial_transformer_block.transformer_blocks[0].ff.net[1].__class__ == nn.Dropout
        assert spatial_transformer_block.transformer_blocks[0].ff.net[2].__class__ == nn.Linear

        dim = 32
        inner_dim = 128

        # First dimension change
        assert spatial_transformer_block.transformer_blocks[0].ff.net[0].proj.in_features == dim
        # NOTE: inner_dim * 2 because GEGLU
        assert spatial_transformer_block.transformer_blocks[0].ff.net[0].proj.out_features == inner_dim * 2

        # Second dimension change
        assert spatial_transformer_block.transformer_blocks[0].ff.net[2].in_features == inner_dim
        assert spatial_transformer_block.transformer_blocks[0].ff.net[2].out_features == dim

    def test_spatial_transformer_geglu_approx_ff_layers(self):
        spatial_transformer_block = Transformer2DModel(
            num_attention_heads=1,
            attention_head_dim=32,
            in_channels=32,
            activation_fn="geglu-approximate",
        )

        assert spatial_transformer_block.transformer_blocks[0].ff.net[0].__class__ == ApproximateGELU
        assert spatial_transformer_block.transformer_blocks[0].ff.net[1].__class__ == nn.Dropout
        assert spatial_transformer_block.transformer_blocks[0].ff.net[2].__class__ == nn.Linear

        dim = 32
        inner_dim = 128

        # First dimension change
        assert spatial_transformer_block.transformer_blocks[0].ff.net[0].proj.in_features == dim
        assert spatial_transformer_block.transformer_blocks[0].ff.net[0].proj.out_features == inner_dim

        # Second dimension change
        assert spatial_transformer_block.transformer_blocks[0].ff.net[2].in_features == inner_dim
        assert spatial_transformer_block.transformer_blocks[0].ff.net[2].out_features == dim

    def test_spatial_transformer_attention_bias(self):
        spatial_transformer_block = Transformer2DModel(
            num_attention_heads=1, attention_head_dim=32, in_channels=32, attention_bias=True
        )

        assert spatial_transformer_block.transformer_blocks[0].attn1.to_q.bias is not None
        assert spatial_transformer_block.transformer_blocks[0].attn1.to_k.bias is not None
        assert spatial_transformer_block.transformer_blocks[0].attn1.to_v.bias is not None