test_stable_diffusion.py 57 KB
Newer Older
1
# coding=utf-8
2
# Copyright 2024 HuggingFace Inc.
3
4
5
6
7
8
9
10
11
12
13
14
15
#
# 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.

16

17
import gc
18
import tempfile
19
import time
20
import traceback
21
22
23
24
import unittest

import numpy as np
import torch
25
from huggingface_hub import hf_hub_download
Aryan's avatar
Aryan committed
26
27
28
29
30
from transformers import (
    CLIPTextConfig,
    CLIPTextModel,
    CLIPTokenizer,
)
31
32
33
34

from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
35
    DPMSolverMultistepScheduler,
hlky's avatar
hlky committed
36
37
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
Patrick von Platen's avatar
Patrick von Platen committed
38
    LCMScheduler,
39
40
41
42
    LMSDiscreteScheduler,
    PNDMScheduler,
    StableDiffusionPipeline,
    UNet2DConditionModel,
43
    logging,
44
)
45
from diffusers.models.attention_processor import AttnProcessor
46
47
48
from diffusers.utils.testing_utils import (
    CaptureLogger,
    enable_full_determinism,
Patrick von Platen's avatar
Patrick von Platen committed
49
    load_image,
Dhruv Nair's avatar
Dhruv Nair committed
50
51
    load_numpy,
    nightly,
52
    numpy_cosine_similarity_distance,
Dhruv Nair's avatar
Dhruv Nair committed
53
    require_python39_or_higher,
54
55
56
    require_torch_2,
    require_torch_gpu,
    run_test_in_subprocess,
Dhruv Nair's avatar
Dhruv Nair committed
57
58
    slow,
    torch_device,
59
)
60

61
62
63
64
65
66
from ..pipeline_params import (
    TEXT_TO_IMAGE_BATCH_PARAMS,
    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
    TEXT_TO_IMAGE_IMAGE_PARAMS,
    TEXT_TO_IMAGE_PARAMS,
)
Aryan's avatar
Aryan committed
67
68
69
70
71
72
from ..test_pipelines_common import (
    IPAdapterTesterMixin,
    PipelineKarrasSchedulerTesterMixin,
    PipelineLatentTesterMixin,
    PipelineTesterMixin,
)
73

74

75
76
77
78
79
80
enable_full_determinism()


# Will be run via run_test_in_subprocess
def _test_stable_diffusion_compile(in_queue, out_queue, timeout):
    error = None
81
    try:
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
        inputs = in_queue.get(timeout=timeout)
        torch_device = inputs.pop("torch_device")
        seed = inputs.pop("seed")
        inputs["generator"] = torch.Generator(device=torch_device).manual_seed(seed)

        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
        sd_pipe.scheduler = DDIMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe = sd_pipe.to(torch_device)

        sd_pipe.unet.to(memory_format=torch.channels_last)
        sd_pipe.unet = torch.compile(sd_pipe.unet, mode="reduce-overhead", fullgraph=True)

        sd_pipe.set_progress_bar_config(disable=None)

        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.38019, 0.28647, 0.27321, 0.40377, 0.38290, 0.35446, 0.39218, 0.38165, 0.42239])
101

102
103
104
105
106
107
108
        assert np.abs(image_slice - expected_slice).max() < 5e-3
    except Exception:
        error = f"{traceback.format_exc()}"

    results = {"error": error}
    out_queue.put(results, timeout=timeout)
    out_queue.join()
109
110


111
class StableDiffusionPipelineFastTests(
Aryan's avatar
Aryan committed
112
113
114
115
116
    IPAdapterTesterMixin,
    PipelineLatentTesterMixin,
    PipelineKarrasSchedulerTesterMixin,
    PipelineTesterMixin,
    unittest.TestCase,
117
):
118
    pipeline_class = StableDiffusionPipeline
119
120
    params = TEXT_TO_IMAGE_PARAMS
    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
121
    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
122
    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
123
    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS
124

Patrick von Platen's avatar
Patrick von Platen committed
125
    def get_dummy_components(self, time_cond_proj_dim=None):
126
        torch.manual_seed(0)
127
        unet = UNet2DConditionModel(
128
129
            block_out_channels=(4, 8),
            layers_per_block=1,
130
            sample_size=32,
Patrick von Platen's avatar
Patrick von Platen committed
131
            time_cond_proj_dim=time_cond_proj_dim,
132
133
134
135
136
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
137
            norm_num_groups=2,
138
        )
139
140
141
142
143
144
        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="scaled_linear",
            clip_sample=False,
            set_alpha_to_one=False,
145
146
        )
        torch.manual_seed(0)
147
        vae = AutoencoderKL(
148
            block_out_channels=[4, 8],
149
150
151
152
153
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=4,
154
            norm_num_groups=2,
155
156
        )
        torch.manual_seed(0)
157
        text_encoder_config = CLIPTextConfig(
158
159
160
            bos_token_id=0,
            eos_token_id=2,
            hidden_size=32,
161
            intermediate_size=64,
162
            layer_norm_eps=1e-05,
163
164
            num_attention_heads=8,
            num_hidden_layers=3,
165
166
167
            pad_token_id=1,
            vocab_size=1000,
        )
168
169
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
170

171
172
173
174
175
176
177
178
        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
            "feature_extractor": None,
179
            "image_encoder": None,
180
181
182
183
184
185
186
187
188
189
190
191
192
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
Aryan's avatar
Aryan committed
193
            "output_type": "np",
194
195
        }
        return inputs
196
197
198
199

    def test_stable_diffusion_ddim(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator

200
201
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
202
        sd_pipe = sd_pipe.to(torch_device)
203
204
        sd_pipe.set_progress_bar_config(disable=None)

205
206
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs)
207
208
209
210
        image = output.images

        image_slice = image[0, -3:, -3:, -1]

211
        assert image.shape == (1, 64, 64, 3)
212
        expected_slice = np.array([0.3203, 0.4555, 0.4711, 0.3505, 0.3973, 0.4650, 0.5137, 0.3392, 0.4045])
213
214
215

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Patrick von Platen's avatar
Patrick von Platen committed
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
    def test_stable_diffusion_lcm(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator

        components = self.get_dummy_components(time_cond_proj_dim=256)
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe.scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs)
        image = output.images

        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.3454, 0.5349, 0.5185, 0.2808, 0.4509, 0.4612, 0.4655, 0.3601, 0.4315])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
    def test_stable_diffusion_lcm_custom_timesteps(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator

        components = self.get_dummy_components(time_cond_proj_dim=256)
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe.scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        del inputs["num_inference_steps"]
        inputs["timesteps"] = [999, 499]
        output = sd_pipe(**inputs)
        image = output.images

        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.3454, 0.5349, 0.5185, 0.2808, 0.4509, 0.4612, 0.4655, 0.3601, 0.4315])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

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
    def test_stable_diffusion_prompt_embeds(self):
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        inputs["prompt"] = 3 * [inputs["prompt"]]

        # forward
        output = sd_pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        inputs = self.get_dummy_inputs(torch_device)
        prompt = 3 * [inputs.pop("prompt")]

        text_inputs = sd_pipe.tokenizer(
            prompt,
            padding="max_length",
            max_length=sd_pipe.tokenizer.model_max_length,
            truncation=True,
            return_tensors="pt",
        )
        text_inputs = text_inputs["input_ids"].to(torch_device)

        prompt_embeds = sd_pipe.text_encoder(text_inputs)[0]

        inputs["prompt_embeds"] = prompt_embeds

        # forward
        output = sd_pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

    def test_stable_diffusion_negative_prompt_embeds(self):
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        negative_prompt = 3 * ["this is a negative prompt"]
        inputs["negative_prompt"] = negative_prompt
        inputs["prompt"] = 3 * [inputs["prompt"]]

        # forward
        output = sd_pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        inputs = self.get_dummy_inputs(torch_device)
        prompt = 3 * [inputs.pop("prompt")]

        embeds = []
        for p in [prompt, negative_prompt]:
            text_inputs = sd_pipe.tokenizer(
                p,
                padding="max_length",
                max_length=sd_pipe.tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_inputs = text_inputs["input_ids"].to(torch_device)

            embeds.append(sd_pipe.text_encoder(text_inputs)[0])

        inputs["prompt_embeds"], inputs["negative_prompt_embeds"] = embeds

        # forward
        output = sd_pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

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
    def test_stable_diffusion_prompt_embeds_with_plain_negative_prompt_list(self):
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        negative_prompt = 3 * ["this is a negative prompt"]
        inputs["negative_prompt"] = negative_prompt
        inputs["prompt"] = 3 * [inputs["prompt"]]

        # forward
        output = sd_pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        inputs = self.get_dummy_inputs(torch_device)
        inputs["negative_prompt"] = negative_prompt
        prompt = 3 * [inputs.pop("prompt")]

        text_inputs = sd_pipe.tokenizer(
            prompt,
            padding="max_length",
            max_length=sd_pipe.tokenizer.model_max_length,
            truncation=True,
            return_tensors="pt",
        )
        text_inputs = text_inputs["input_ids"].to(torch_device)

        prompt_embeds = sd_pipe.text_encoder(text_inputs)[0]

        inputs["prompt_embeds"] = prompt_embeds

        # forward
        output = sd_pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

373
374
375
376
377
378
    def test_ip_adapter_single(self):
        expected_pipe_slice = None
        if torch_device == "cpu":
            expected_pipe_slice = np.array([0.3203, 0.4555, 0.4711, 0.3505, 0.3973, 0.4650, 0.5137, 0.3392, 0.4045])
        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)

379
380
381
    def test_stable_diffusion_ddim_factor_8(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator

382
383
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
384
385
386
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

387
388
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs, height=136, width=136)
389
390
391
392
        image = output.images

        image_slice = image[0, -3:, -3:, -1]

393
        assert image.shape == (1, 136, 136, 3)
394
        expected_slice = np.array([0.4346, 0.5621, 0.5016, 0.3926, 0.4533, 0.4134, 0.5625, 0.5632, 0.5265])
395
396
397
398
399

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_pndm(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
400
401
402
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe.scheduler = PNDMScheduler(skip_prk_steps=True)
403
404
405
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

406
407
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs)
408
409
410
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

411
        assert image.shape == (1, 64, 64, 3)
412
        expected_slice = np.array([0.3411, 0.5032, 0.4704, 0.3135, 0.4323, 0.4740, 0.5150, 0.3498, 0.4022])
413

414
415
416
417
418
419
420
421
422
423
424
425
426
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_no_safety_checker(self):
        pipe = StableDiffusionPipeline.from_pretrained(
            "hf-internal-testing/tiny-stable-diffusion-lms-pipe", safety_checker=None
        )
        assert isinstance(pipe, StableDiffusionPipeline)
        assert isinstance(pipe.scheduler, LMSDiscreteScheduler)
        assert pipe.safety_checker is None

        image = pipe("example prompt", num_inference_steps=2).images[0]
        assert image is not None

427
428
429
430
431
432
433
434
435
436
        # check that there's no error when saving a pipeline with one of the models being None
        with tempfile.TemporaryDirectory() as tmpdirname:
            pipe.save_pretrained(tmpdirname)
            pipe = StableDiffusionPipeline.from_pretrained(tmpdirname)

        # sanity check that the pipeline still works
        assert pipe.safety_checker is None
        image = pipe("example prompt", num_inference_steps=2).images[0]
        assert image is not None

437
438
    def test_stable_diffusion_k_lms(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
hlky's avatar
hlky committed
439

440
441
442
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config)
hlky's avatar
hlky committed
443
444
445
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

446
447
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs)
hlky's avatar
hlky committed
448
449
450
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

451
        assert image.shape == (1, 64, 64, 3)
452
        expected_slice = np.array([0.3149, 0.5246, 0.4796, 0.3218, 0.4469, 0.4729, 0.5151, 0.3597, 0.3954])
453

hlky's avatar
hlky committed
454
455
456
457
458
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_k_euler_ancestral(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator

459
460
461
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(sd_pipe.scheduler.config)
hlky's avatar
hlky committed
462
463
464
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

465
466
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs)
hlky's avatar
hlky committed
467
468
469
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

470
        assert image.shape == (1, 64, 64, 3)
471
        expected_slice = np.array([0.3151, 0.5243, 0.4794, 0.3217, 0.4468, 0.4728, 0.5152, 0.3598, 0.3954])
472

hlky's avatar
hlky committed
473
474
475
476
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_k_euler(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
477

478
479
480
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe.scheduler = EulerDiscreteScheduler.from_config(sd_pipe.scheduler.config)
481
482
483
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

484
485
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs)
486
487
488
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

489
        assert image.shape == (1, 64, 64, 3)
490
        expected_slice = np.array([0.3149, 0.5246, 0.4796, 0.3218, 0.4469, 0.4729, 0.5151, 0.3597, 0.3954])
491

492
493
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

494
495
    def test_stable_diffusion_vae_slicing(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
496
497
498
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
499
500
501
502
503
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        image_count = 4

504
505
506
        inputs = self.get_dummy_inputs(device)
        inputs["prompt"] = [inputs["prompt"]] * image_count
        output_1 = sd_pipe(**inputs)
507
508
509

        # make sure sliced vae decode yields the same result
        sd_pipe.enable_vae_slicing()
510
511
512
        inputs = self.get_dummy_inputs(device)
        inputs["prompt"] = [inputs["prompt"]] * image_count
        output_2 = sd_pipe(**inputs)
513
514
515
516

        # there is a small discrepancy at image borders vs. full batch decode
        assert np.abs(output_2.images.flatten() - output_1.images.flatten()).max() < 3e-3

517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
    def test_stable_diffusion_vae_tiling(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()

        # make sure here that pndm scheduler skips prk
        components["safety_checker"] = None
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"

        # Test that tiled decode at 512x512 yields the same result as the non-tiled decode
        generator = torch.Generator(device=device).manual_seed(0)
        output_1 = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")

        # make sure tiled vae decode yields the same result
        sd_pipe.enable_vae_tiling()
        generator = torch.Generator(device=device).manual_seed(0)
        output_2 = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")

        assert np.abs(output_2.images.flatten() - output_1.images.flatten()).max() < 5e-1

540
541
542
543
544
545
        # test that tiled decode works with various shapes
        shapes = [(1, 4, 73, 97), (1, 4, 97, 73), (1, 4, 49, 65), (1, 4, 65, 49)]
        for shape in shapes:
            zeros = torch.zeros(shape).to(device)
            sd_pipe.vae.decode(zeros)

546
547
    def test_stable_diffusion_negative_prompt(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
548
549
550
        components = self.get_dummy_components()
        components["scheduler"] = PNDMScheduler(skip_prk_steps=True)
        sd_pipe = StableDiffusionPipeline(**components)
551
552
553
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

554
        inputs = self.get_dummy_inputs(device)
555
        negative_prompt = "french fries"
556
        output = sd_pipe(**inputs, negative_prompt=negative_prompt)
557
558
559
560

        image = output.images
        image_slice = image[0, -3:, -3:, -1]

561
        assert image.shape == (1, 64, 64, 3)
562
        expected_slice = np.array([0.3458, 0.5120, 0.4800, 0.3116, 0.4348, 0.4802, 0.5237, 0.3467, 0.3991])
563

564
565
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

566
    def test_stable_diffusion_long_prompt(self):
567
568
569
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
570
571
572
573
574
575
576
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        do_classifier_free_guidance = True
        negative_prompt = None
        num_images_per_prompt = 1
        logger = logging.get_logger("diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion")
577
        logger.setLevel(logging.WARNING)
578
579
580

        prompt = 100 * "@"
        with CaptureLogger(logger) as cap_logger:
581
            negative_text_embeddings, text_embeddings = sd_pipe.encode_prompt(
582
583
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )
584
585
            if negative_text_embeddings is not None:
                text_embeddings = torch.cat([negative_text_embeddings, text_embeddings])
586

587
588
589
        # 100 - 77 + 1 (BOS token) + 1 (EOS token) = 25
        assert cap_logger.out.count("@") == 25

590
591
        negative_prompt = "Hello"
        with CaptureLogger(logger) as cap_logger_2:
592
            negative_text_embeddings_2, text_embeddings_2 = sd_pipe.encode_prompt(
593
594
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )
595
596
            if negative_text_embeddings_2 is not None:
                text_embeddings_2 = torch.cat([negative_text_embeddings_2, text_embeddings_2])
597

598
599
600
601
602
603
604
605
606
607
        assert cap_logger.out == cap_logger_2.out

        prompt = 25 * "@"
        with CaptureLogger(logger) as cap_logger_3:
            negative_text_embeddings_3, text_embeddings_3 = sd_pipe.encode_prompt(
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )
            if negative_text_embeddings_3 is not None:
                text_embeddings_3 = torch.cat([negative_text_embeddings_3, text_embeddings_3])

608
609
610
611
        assert text_embeddings_3.shape == text_embeddings_2.shape == text_embeddings.shape
        assert text_embeddings.shape[1] == 77
        assert cap_logger_3.out == ""

612
    def test_stable_diffusion_height_width_opt(self):
613
614
615
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
616
617
618
619
620
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "hey"

621
        output = sd_pipe(prompt, num_inference_steps=1, output_type="np")
622
        image_shape = output.images[0].shape[:2]
Patrick von Platen's avatar
Patrick von Platen committed
623
        assert image_shape == (64, 64)
624

625
        output = sd_pipe(prompt, num_inference_steps=1, height=96, width=96, output_type="np")
626
        image_shape = output.images[0].shape[:2]
Patrick von Platen's avatar
Patrick von Platen committed
627
        assert image_shape == (96, 96)
628
629
630

        config = dict(sd_pipe.unet.config)
        config["sample_size"] = 96
Patrick von Platen's avatar
Patrick von Platen committed
631
        sd_pipe.unet = UNet2DConditionModel.from_config(config).to(torch_device)
632
        output = sd_pipe(prompt, num_inference_steps=1, output_type="np")
633
        image_shape = output.images[0].shape[:2]
Patrick von Platen's avatar
Patrick von Platen committed
634
        assert image_shape == (192, 192)
635

636
637
638
639
640
641
    def test_attention_slicing_forward_pass(self):
        super().test_attention_slicing_forward_pass(expected_max_diff=3e-3)

    def test_inference_batch_single_identical(self):
        super().test_inference_batch_single_identical(expected_max_diff=3e-3)

642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
    def test_freeu_enabled(self):
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "hey"
        output = sd_pipe(prompt, num_inference_steps=1, output_type="np", generator=torch.manual_seed(0)).images

        sd_pipe.enable_freeu(s1=0.9, s2=0.2, b1=1.2, b2=1.4)
        output_freeu = sd_pipe(prompt, num_inference_steps=1, output_type="np", generator=torch.manual_seed(0)).images

        assert not np.allclose(
            output[0, -3:, -3:, -1], output_freeu[0, -3:, -3:, -1]
        ), "Enabling of FreeU should lead to different results."

    def test_freeu_disabled(self):
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "hey"
        output = sd_pipe(prompt, num_inference_steps=1, output_type="np", generator=torch.manual_seed(0)).images

        sd_pipe.enable_freeu(s1=0.9, s2=0.2, b1=1.2, b2=1.4)
        sd_pipe.disable_freeu()

        freeu_keys = {"s1", "s2", "b1", "b2"}
        for upsample_block in sd_pipe.unet.up_blocks:
            for key in freeu_keys:
                assert getattr(upsample_block, key) is None, f"Disabling of FreeU should have set {key} to None."

        output_no_freeu = sd_pipe(
            prompt, num_inference_steps=1, output_type="np", generator=torch.manual_seed(0)
        ).images

        assert np.allclose(
            output[0, -3:, -3:, -1], output_no_freeu[0, -3:, -3:, -1]
        ), "Disabling of FreeU should lead to results similar to the default pipeline results."

683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
    def test_fused_qkv_projections(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        original_image_slice = image[0, -3:, -3:, -1]

        sd_pipe.fuse_qkv_projections()
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice_fused = image[0, -3:, -3:, -1]

        sd_pipe.unfuse_qkv_projections()
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice_disabled = image[0, -3:, -3:, -1]

        assert np.allclose(
            original_image_slice, image_slice_fused, atol=1e-2, rtol=1e-2
        ), "Fusion of QKV projections shouldn't affect the outputs."
        assert np.allclose(
            image_slice_fused, image_slice_disabled, atol=1e-2, rtol=1e-2
        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
        assert np.allclose(
            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
        ), "Original outputs should match when fused QKV projections are disabled."

Dhruv Nair's avatar
Dhruv Nair committed
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
    def test_pipeline_interrupt(self):
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "hey"
        num_inference_steps = 3

        # store intermediate latents from the generation process
        class PipelineState:
            def __init__(self):
                self.state = []

            def apply(self, pipe, i, t, callback_kwargs):
                self.state.append(callback_kwargs["latents"])
                return callback_kwargs

        pipe_state = PipelineState()
        sd_pipe(
            prompt,
            num_inference_steps=num_inference_steps,
            output_type="np",
            generator=torch.Generator("cpu").manual_seed(0),
            callback_on_step_end=pipe_state.apply,
        ).images

        # interrupt generation at step index
        interrupt_step_idx = 1

        def callback_on_step_end(pipe, i, t, callback_kwargs):
            if i == interrupt_step_idx:
                pipe._interrupt = True

            return callback_kwargs

        output_interrupted = sd_pipe(
            prompt,
            num_inference_steps=num_inference_steps,
            output_type="latent",
            generator=torch.Generator("cpu").manual_seed(0),
            callback_on_step_end=callback_on_step_end,
        ).images

        # fetch intermediate latents at the interrupted step
        # from the completed generation process
        intermediate_latent = pipe_state.state[interrupt_step_idx]

        # compare the intermediate latent to the output of the interrupted process
        # they should be the same
        assert torch.allclose(intermediate_latent, output_interrupted, atol=1e-4)

766
767

@slow
768
@require_torch_gpu
769
class StableDiffusionPipelineSlowTests(unittest.TestCase):
770
    def setUp(self):
771
772
773
        gc.collect()
        torch.cuda.empty_cache()

774
775
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
776
777
778
779
780
781
782
783
        latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
        inputs = {
            "prompt": "a photograph of an astronaut riding a horse",
            "latents": latents,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 7.5,
784
            "output_type": "np",
785
786
787
788
        }
        return inputs

    def test_stable_diffusion_1_1_pndm(self):
789
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-1")
790
791
792
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

793
794
795
        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
796

797
        assert image.shape == (1, 512, 512, 3)
Dhruv Nair's avatar
Dhruv Nair committed
798
        expected_slice = np.array([0.4363, 0.4355, 0.3667, 0.4066, 0.3970, 0.3866, 0.4394, 0.4356, 0.4059])
799
        assert np.abs(image_slice - expected_slice).max() < 3e-3
800

801
802
803
804
805
806
807
808
809
810
811
812
813
814
    def test_stable_diffusion_v1_4_with_freeu(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4").to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 25

        sd_pipe.enable_freeu(s1=0.9, s2=0.2, b1=1.2, b2=1.4)
        image = sd_pipe(**inputs).images
        image = image[0, -3:, -3:, -1].flatten()
        expected_image = [0.0721, 0.0588, 0.0268, 0.0384, 0.0636, 0.0, 0.0429, 0.0344, 0.0309]
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

815
816
817
818
    def test_stable_diffusion_1_4_pndm(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)
819

820
821
822
        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
823

824
        assert image.shape == (1, 512, 512, 3)
Dhruv Nair's avatar
Dhruv Nair committed
825
        expected_slice = np.array([0.5740, 0.4784, 0.3162, 0.6358, 0.5831, 0.5505, 0.5082, 0.5631, 0.5575])
826
        assert np.abs(image_slice - expected_slice).max() < 3e-3
827

828
829
830
    def test_stable_diffusion_ddim(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
        sd_pipe.scheduler = DDIMScheduler.from_config(sd_pipe.scheduler.config)
831
832
833
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

834
835
836
        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
837

838
839
840
        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.38019, 0.28647, 0.27321, 0.40377, 0.38290, 0.35446, 0.39218, 0.38165, 0.42239])
        assert np.abs(image_slice - expected_slice).max() < 1e-4
841

842
843
844
845
846
847
848
849
850
    def test_stable_diffusion_lms(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
        sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
851
852

        assert image.shape == (1, 512, 512, 3)
853
        expected_slice = np.array([0.10542, 0.09620, 0.07332, 0.09015, 0.09382, 0.07597, 0.08496, 0.07806, 0.06455])
854
        assert np.abs(image_slice - expected_slice).max() < 3e-3
855

856
857
    def test_stable_diffusion_dpm(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
858
859
860
861
        sd_pipe.scheduler = DPMSolverMultistepScheduler.from_config(
            sd_pipe.scheduler.config,
            final_sigmas_type="sigma_min",
        )
862
863
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)
864

865
866
867
        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
868
869

        assert image.shape == (1, 512, 512, 3)
870
        expected_slice = np.array([0.03503, 0.03494, 0.01087, 0.03128, 0.02552, 0.00803, 0.00742, 0.00372, 0.00000])
871
        assert np.abs(image_slice - expected_slice).max() < 3e-3
872

873
    def test_stable_diffusion_attention_slicing(self):
874
        torch.cuda.reset_peak_memory_stats()
875
        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16)
876
        pipe.unet.set_default_attn_processor()
877
        pipe = pipe.to(torch_device)
878
879
        pipe.set_progress_bar_config(disable=None)

880
        # enable attention slicing
881
        pipe.enable_attention_slicing()
882
883
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image_sliced = pipe(**inputs).images
884
885
886
887
888
889

        mem_bytes = torch.cuda.max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()
        # make sure that less than 3.75 GB is allocated
        assert mem_bytes < 3.75 * 10**9

890
        # disable slicing
891
        pipe.disable_attention_slicing()
892
        pipe.unet.set_default_attn_processor()
893
894
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image = pipe(**inputs).images
895
896
897
898

        # make sure that more than 3.75 GB is allocated
        mem_bytes = torch.cuda.max_memory_allocated()
        assert mem_bytes > 3.75 * 10**9
899
900
        max_diff = numpy_cosine_similarity_distance(image_sliced.flatten(), image.flatten())
        assert max_diff < 1e-3
901

902
903
    def test_stable_diffusion_vae_slicing(self):
        torch.cuda.reset_peak_memory_stats()
904
        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16)
905
        pipe = pipe.to(torch_device)
906
907
908
909
910
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        # enable vae slicing
        pipe.enable_vae_slicing()
911
912
913
914
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        inputs["prompt"] = [inputs["prompt"]] * 4
        inputs["latents"] = torch.cat([inputs["latents"]] * 4)
        image_sliced = pipe(**inputs).images
915
916
917
918
919
920
921
922

        mem_bytes = torch.cuda.max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()
        # make sure that less than 4 GB is allocated
        assert mem_bytes < 4e9

        # disable vae slicing
        pipe.disable_vae_slicing()
923
924
925
926
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        inputs["prompt"] = [inputs["prompt"]] * 4
        inputs["latents"] = torch.cat([inputs["latents"]] * 4)
        image = pipe(**inputs).images
927
928
929
930
931

        # make sure that more than 4 GB is allocated
        mem_bytes = torch.cuda.max_memory_allocated()
        assert mem_bytes > 4e9
        # There is a small discrepancy at the image borders vs. a fully batched version.
932
933
        max_diff = numpy_cosine_similarity_distance(image_sliced.flatten(), image.flatten())
        assert max_diff < 1e-2
934

935
936
937
    def test_stable_diffusion_vae_tiling(self):
        torch.cuda.reset_peak_memory_stats()
        model_id = "CompVis/stable-diffusion-v1-4"
938
939
940
        pipe = StableDiffusionPipeline.from_pretrained(
            model_id, revision="fp16", torch_dtype=torch.float16, safety_checker=None
        )
941
942
943
944
945
946
947
948
949
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()
        pipe.unet = pipe.unet.to(memory_format=torch.channels_last)
        pipe.vae = pipe.vae.to(memory_format=torch.channels_last)

        prompt = "a photograph of an astronaut riding a horse"

        # enable vae tiling
        pipe.enable_vae_tiling()
950
951
952
953
954
955
956
957
958
        pipe.enable_model_cpu_offload()
        generator = torch.Generator(device="cpu").manual_seed(0)
        output_chunked = pipe(
            [prompt],
            width=1024,
            height=1024,
            generator=generator,
            guidance_scale=7.5,
            num_inference_steps=2,
959
            output_type="np",
960
961
        )
        image_chunked = output_chunked.images
962
963
964
965
966

        mem_bytes = torch.cuda.max_memory_allocated()

        # disable vae tiling
        pipe.disable_vae_tiling()
967
968
969
970
971
972
973
974
        generator = torch.Generator(device="cpu").manual_seed(0)
        output = pipe(
            [prompt],
            width=1024,
            height=1024,
            generator=generator,
            guidance_scale=7.5,
            num_inference_steps=2,
975
            output_type="np",
976
977
        )
        image = output.images
978

979
        assert mem_bytes < 1e10
980
981
        max_diff = numpy_cosine_similarity_distance(image_chunked.flatten(), image.flatten())
        assert max_diff < 1e-2
982

983
    def test_stable_diffusion_fp16_vs_autocast(self):
984
985
        # this test makes sure that the original model with autocast
        # and the new model with fp16 yield the same result
986
        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16)
987
        pipe = pipe.to(torch_device)
988
989
        pipe.set_progress_bar_config(disable=None)

990
991
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image_fp16 = pipe(**inputs).images
992
993

        with torch.autocast(torch_device):
994
995
            inputs = self.get_inputs(torch_device)
            image_autocast = pipe(**inputs).images
996
997

        # Make sure results are close enough
998
        diff = np.abs(image_fp16.flatten() - image_autocast.flatten())
999
1000
1001
1002
        # They ARE different since ops are not run always at the same precision
        # however, they should be extremely close.
        assert diff.mean() < 2e-2

1003
    def test_stable_diffusion_intermediate_state(self):
1004
1005
        number_of_steps = 0

1006
1007
        def callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None:
            callback_fn.has_been_called = True
1008
1009
            nonlocal number_of_steps
            number_of_steps += 1
1010
            if step == 1:
1011
1012
1013
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
1014
1015
1016
1017
1018
                expected_slice = np.array(
                    [-0.5693, -0.3018, -0.9746, 0.0518, -0.8770, 0.7559, -1.7402, 0.1022, 1.1582]
                )

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
1019
            elif step == 2:
1020
1021
1022
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
1023
1024
1025
1026
1027
                expected_slice = np.array(
                    [-0.1958, -0.2993, -1.0166, -0.5005, -0.4810, 0.6162, -0.9492, 0.6621, 1.4492]
                )

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
1028

1029
        callback_fn.has_been_called = False
1030

1031
        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16)
1032
1033
1034
1035
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

1036
1037
1038
1039
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        pipe(**inputs, callback=callback_fn, callback_steps=1)
        assert callback_fn.has_been_called
        assert number_of_steps == inputs["num_inference_steps"]
1040

1041
    def test_stable_diffusion_low_cpu_mem_usage(self):
1042
1043
1044
        pipeline_id = "CompVis/stable-diffusion-v1-4"

        start_time = time.time()
1045
        pipeline_low_cpu_mem_usage = StableDiffusionPipeline.from_pretrained(pipeline_id, torch_dtype=torch.float16)
1046
1047
        pipeline_low_cpu_mem_usage.to(torch_device)
        low_cpu_mem_usage_time = time.time() - start_time
1048
1049

        start_time = time.time()
1050
        _ = StableDiffusionPipeline.from_pretrained(pipeline_id, torch_dtype=torch.float16, low_cpu_mem_usage=False)
1051
        normal_load_time = time.time() - start_time
1052

1053
        assert 2 * low_cpu_mem_usage_time < normal_load_time
1054

1055
    def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self):
1056
1057
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
Anton Lozhkov's avatar
Anton Lozhkov committed
1058
        torch.cuda.reset_peak_memory_stats()
1059

1060
        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16)
1061
1062
1063
1064
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()
1065

1066
1067
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = pipe(**inputs)
1068
1069

        mem_bytes = torch.cuda.max_memory_allocated()
Anton Lozhkov's avatar
Anton Lozhkov committed
1070
1071
        # make sure that less than 2.8 GB is allocated
        assert mem_bytes < 2.8 * 10**9
1072

1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
    def test_stable_diffusion_pipeline_with_model_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)

        # Normal inference

        pipe = StableDiffusionPipeline.from_pretrained(
            "CompVis/stable-diffusion-v1-4",
            torch_dtype=torch.float16,
        )
1086
        pipe.unet.set_default_attn_processor()
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        outputs = pipe(**inputs)
        mem_bytes = torch.cuda.max_memory_allocated()

        # With model offloading

        # Reload but don't move to cuda
        pipe = StableDiffusionPipeline.from_pretrained(
            "CompVis/stable-diffusion-v1-4",
            torch_dtype=torch.float16,
        )
1099
        pipe.unet.set_default_attn_processor()
1100
1101
1102
1103
1104
1105
1106

        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe.enable_model_cpu_offload()
        pipe.set_progress_bar_config(disable=None)
1107
1108
        inputs = self.get_inputs(torch_device, dtype=torch.float16)

1109
1110
1111
        outputs_offloaded = pipe(**inputs)
        mem_bytes_offloaded = torch.cuda.max_memory_allocated()

1112
1113
1114
1115
1116
        images = outputs.images
        offloaded_images = outputs_offloaded.images

        max_diff = numpy_cosine_similarity_distance(images.flatten(), offloaded_images.flatten())
        assert max_diff < 1e-3
1117
1118
        assert mem_bytes_offloaded < mem_bytes
        assert mem_bytes_offloaded < 3.5 * 10**9
1119
        for module in pipe.text_encoder, pipe.unet, pipe.vae:
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
            assert module.device == torch.device("cpu")

        # With attention slicing
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe.enable_attention_slicing()
        _ = pipe(**inputs)
        mem_bytes_slicing = torch.cuda.max_memory_allocated()

        assert mem_bytes_slicing < mem_bytes_offloaded
        assert mem_bytes_slicing < 3 * 10**9

1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
    def test_stable_diffusion_textual_inversion(self):
        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
        pipe.load_textual_inversion("sd-concepts-library/low-poly-hd-logos-icons")

        a111_file = hf_hub_download("hf-internal-testing/text_inv_embedding_a1111_format", "winter_style.pt")
        a111_file_neg = hf_hub_download(
            "hf-internal-testing/text_inv_embedding_a1111_format", "winter_style_negative.pt"
        )
        pipe.load_textual_inversion(a111_file)
        pipe.load_textual_inversion(a111_file_neg)
        pipe.to("cuda")

        generator = torch.Generator(device="cpu").manual_seed(1)

        prompt = "An logo of a turtle in strong Style-Winter with <low-poly-hd-logos-icons>"
        neg_prompt = "Style-Winter-neg"

        image = pipe(prompt=prompt, negative_prompt=neg_prompt, generator=generator, output_type="np").images[0]
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/text_inv/winter_logo_style.npy"
        )

        max_diff = np.abs(expected_image - image).max()
1157
        assert max_diff < 8e-1
1158

1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
    def test_stable_diffusion_textual_inversion_with_model_cpu_offload(self):
        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
        pipe.enable_model_cpu_offload()
        pipe.load_textual_inversion("sd-concepts-library/low-poly-hd-logos-icons")

        a111_file = hf_hub_download("hf-internal-testing/text_inv_embedding_a1111_format", "winter_style.pt")
        a111_file_neg = hf_hub_download(
            "hf-internal-testing/text_inv_embedding_a1111_format", "winter_style_negative.pt"
        )
        pipe.load_textual_inversion(a111_file)
        pipe.load_textual_inversion(a111_file_neg)

        generator = torch.Generator(device="cpu").manual_seed(1)

        prompt = "An logo of a turtle in strong Style-Winter with <low-poly-hd-logos-icons>"
        neg_prompt = "Style-Winter-neg"

        image = pipe(prompt=prompt, negative_prompt=neg_prompt, generator=generator, output_type="np").images[0]
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/text_inv/winter_logo_style.npy"
        )

        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 8e-1

    def test_stable_diffusion_textual_inversion_with_sequential_cpu_offload(self):
        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
        pipe.enable_sequential_cpu_offload()
        pipe.load_textual_inversion("sd-concepts-library/low-poly-hd-logos-icons")

        a111_file = hf_hub_download("hf-internal-testing/text_inv_embedding_a1111_format", "winter_style.pt")
        a111_file_neg = hf_hub_download(
            "hf-internal-testing/text_inv_embedding_a1111_format", "winter_style_negative.pt"
        )
        pipe.load_textual_inversion(a111_file)
        pipe.load_textual_inversion(a111_file_neg)

        generator = torch.Generator(device="cpu").manual_seed(1)

        prompt = "An logo of a turtle in strong Style-Winter with <low-poly-hd-logos-icons>"
        neg_prompt = "Style-Winter-neg"

        image = pipe(prompt=prompt, negative_prompt=neg_prompt, generator=generator, output_type="np").images[0]
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/text_inv/winter_logo_style.npy"
        )

        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 8e-1

Dhruv Nair's avatar
Dhruv Nair committed
1209
    @require_python39_or_higher
1210
    @require_torch_2
1211
    def test_stable_diffusion_compile(self):
1212
1213
1214
1215
1216
1217
1218
        seed = 0
        inputs = self.get_inputs(torch_device, seed=seed)
        # Can't pickle a Generator object
        del inputs["generator"]
        inputs["torch_device"] = torch_device
        inputs["seed"] = seed
        run_test_in_subprocess(test_case=self, target_func=_test_stable_diffusion_compile, inputs=inputs)
1219

Patrick von Platen's avatar
Patrick von Platen committed
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
    def test_stable_diffusion_lcm(self):
        unet = UNet2DConditionModel.from_pretrained("SimianLuo/LCM_Dreamshaper_v7", subfolder="unet")
        sd_pipe = StableDiffusionPipeline.from_pretrained("Lykon/dreamshaper-7", unet=unet).to(torch_device)
        sd_pipe.scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 6
        inputs["output_type"] = "pil"

        image = sd_pipe(**inputs).images[0]

        expected_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/lcm_full/stable_diffusion_lcm.png"
        )

        image = sd_pipe.image_processor.pil_to_numpy(image)
        expected_image = sd_pipe.image_processor.pil_to_numpy(expected_image)

        max_diff = numpy_cosine_similarity_distance(image.flatten(), expected_image.flatten())

        assert max_diff < 1e-2

1243

1lint's avatar
1lint committed
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
@slow
@require_torch_gpu
class StableDiffusionPipelineCkptTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def test_download_from_hub(self):
        ckpt_paths = [
            "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt",
            "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix_base.ckpt",
        ]

        for ckpt_path in ckpt_paths:
Patrick von Platen's avatar
Patrick von Platen committed
1259
            pipe = StableDiffusionPipeline.from_single_file(ckpt_path, torch_dtype=torch.float16)
1lint's avatar
1lint committed
1260
1261
1262
1263
1264
1265
1266
1267
            pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
            pipe.to("cuda")

        image_out = pipe("test", num_inference_steps=1, output_type="np").images[0]

        assert image_out.shape == (512, 512, 3)

    def test_download_local(self):
1268
1269
        ckpt_filename = hf_hub_download("runwayml/stable-diffusion-v1-5", filename="v1-5-pruned-emaonly.ckpt")
        config_filename = hf_hub_download("runwayml/stable-diffusion-v1-5", filename="v1-inference.yaml")
1lint's avatar
1lint committed
1270

1271
1272
1273
        pipe = StableDiffusionPipeline.from_single_file(
            ckpt_filename, config_files={"v1": config_filename}, torch_dtype=torch.float16
        )
1lint's avatar
1lint committed
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.to("cuda")

        image_out = pipe("test", num_inference_steps=1, output_type="np").images[0]

        assert image_out.shape == (512, 512, 3)

    def test_download_ckpt_diff_format_is_same(self):
        ckpt_path = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt"

1284
1285
1286
1287
        sf_pipe = StableDiffusionPipeline.from_single_file(ckpt_path)
        sf_pipe.scheduler = DDIMScheduler.from_config(sf_pipe.scheduler.config)
        sf_pipe.unet.set_attn_processor(AttnProcessor())
        sf_pipe.to("cuda")
1lint's avatar
1lint committed
1288
1289

        generator = torch.Generator(device="cpu").manual_seed(0)
1290
        image_single_file = sf_pipe("a turtle", num_inference_steps=2, generator=generator, output_type="np").images[0]
1lint's avatar
1lint committed
1291
1292
1293
1294
1295
1296
1297

        pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.unet.set_attn_processor(AttnProcessor())
        pipe.to("cuda")

        generator = torch.Generator(device="cpu").manual_seed(0)
1298
        image = pipe("a turtle", num_inference_steps=2, generator=generator, output_type="np").images[0]
1lint's avatar
1lint committed
1299

1300
        max_diff = numpy_cosine_similarity_distance(image.flatten(), image_single_file.flatten())
1301
1302

        assert max_diff < 1e-3
1lint's avatar
1lint committed
1303

1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
    def test_single_file_component_configs(self):
        pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")

        ckpt_path = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt"
        single_file_pipe = StableDiffusionPipeline.from_single_file(ckpt_path, load_safety_checker=True)

        for param_name, param_value in single_file_pipe.text_encoder.config.to_dict().items():
            if param_name in ["torch_dtype", "architectures", "_name_or_path"]:
                continue
            assert pipe.text_encoder.config.to_dict()[param_name] == param_value

        PARAMS_TO_IGNORE = ["torch_dtype", "_name_or_path", "architectures", "_use_default_values"]
        for param_name, param_value in single_file_pipe.unet.config.items():
            if param_name in PARAMS_TO_IGNORE:
                continue
            assert (
                pipe.unet.config[param_name] == param_value
            ), f"{param_name} differs between single file loading and pretrained loading"

        for param_name, param_value in single_file_pipe.vae.config.items():
            if param_name in PARAMS_TO_IGNORE:
                continue
            assert (
                pipe.vae.config[param_name] == param_value
            ), f"{param_name} differs between single file loading and pretrained loading"

        for param_name, param_value in single_file_pipe.safety_checker.config.to_dict().items():
            if param_name in PARAMS_TO_IGNORE:
                continue
            assert (
                pipe.safety_checker.config.to_dict()[param_name] == param_value
            ), f"{param_name} differs between single file loading and pretrained loading"

1lint's avatar
1lint committed
1337

1338
1339
1340
1341
1342
1343
1344
1345
@nightly
@require_torch_gpu
class StableDiffusionPipelineNightlyTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

1346
1347
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
1348
1349
1350
1351
1352
1353
1354
1355
        latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
        inputs = {
            "prompt": "a photograph of an astronaut riding a horse",
            "latents": latents,
            "generator": generator,
            "num_inference_steps": 50,
            "guidance_scale": 7.5,
1356
            "output_type": "np",
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
        }
        return inputs

    def test_stable_diffusion_1_4_pndm(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4").to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_text2img/stable_diffusion_1_4_pndm.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_1_5_pndm(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5").to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_text2img/stable_diffusion_1_5_pndm.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_ddim(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4").to(torch_device)
        sd_pipe.scheduler = DDIMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_text2img/stable_diffusion_1_4_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
1401
        assert max_diff < 3e-3
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431

    def test_stable_diffusion_lms(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4").to(torch_device)
        sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_text2img/stable_diffusion_1_4_lms.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_euler(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4").to(torch_device)
        sd_pipe.scheduler = EulerDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_text2img/stable_diffusion_1_4_euler.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3