test_pipelines_common.py 53.1 KB
Newer Older
1
2
3
4
import contextlib
import gc
import inspect
import io
5
6
import json
import os
7
8
9
import re
import tempfile
import unittest
10
import uuid
11
12
13
from typing import Callable, Union

import numpy as np
Anh71me's avatar
Anh71me committed
14
import PIL.Image
15
import torch
16
17
from huggingface_hub import delete_repo
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
18

19
import diffusers
20
21
22
23
24
25
26
27
28
29
from diffusers import (
    AsymmetricAutoencoderKL,
    AutoencoderKL,
    AutoencoderTiny,
    ConsistencyDecoderVAE,
    DDIMScheduler,
    DiffusionPipeline,
    StableDiffusionPipeline,
    UNet2DConditionModel,
)
30
from diffusers.image_processor import VaeImageProcessor
31
from diffusers.schedulers import KarrasDiffusionSchedulers
32
from diffusers.utils import logging
33
from diffusers.utils.import_utils import is_accelerate_available, is_accelerate_version, is_xformers_available
34
35
36
37
38
from diffusers.utils.testing_utils import (
    CaptureLogger,
    require_torch,
    torch_device,
)
39

40
41
42
43
44
45
from ..models.test_models_vae import (
    get_asym_autoencoder_kl_config,
    get_autoencoder_kl_config,
    get_autoencoder_tiny_config,
    get_consistency_vae_config,
)
46
47
from ..others.test_utils import TOKEN, USER, is_staging_test

48

49
50
51
52
53
54
55
def to_np(tensor):
    if isinstance(tensor, torch.Tensor):
        tensor = tensor.detach().cpu().numpy()

    return tensor


56
57
58
59
60
def check_same_shape(tensor_list):
    shapes = [tensor.shape for tensor in tensor_list]
    return all(shape == shapes[0] for shape in shapes[1:])


61
62
63
64
65
66
67
68
69
70
71
72
73
74
class PipelineLatentTesterMixin:
    """
    This mixin is designed to be used with PipelineTesterMixin and unittest.TestCase classes.
    It provides a set of common tests for PyTorch pipeline that has vae, e.g.
    equivalence of different input and output types, etc.
    """

    @property
    def image_params(self) -> frozenset:
        raise NotImplementedError(
            "You need to set the attribute `image_params` in the child test class. "
            "`image_params` are tested for if all accepted input image types (i.e. `pt`,`pil`,`np`) are producing same results"
        )

75
76
77
78
79
80
81
    @property
    def image_latents_params(self) -> frozenset:
        raise NotImplementedError(
            "You need to set the attribute `image_latents_params` in the child test class. "
            "`image_latents_params` are tested for if passing latents directly are producing same results"
        )

82
83
84
    def get_dummy_inputs_by_type(self, device, seed=0, input_image_type="pt", output_type="np"):
        inputs = self.get_dummy_inputs(device, seed)

85
86
87
88
89
90
91
92
93
94
95
96
        def convert_to_pt(image):
            if isinstance(image, torch.Tensor):
                input_image = image
            elif isinstance(image, np.ndarray):
                input_image = VaeImageProcessor.numpy_to_pt(image)
            elif isinstance(image, PIL.Image.Image):
                input_image = VaeImageProcessor.pil_to_numpy(image)
                input_image = VaeImageProcessor.numpy_to_pt(input_image)
            else:
                raise ValueError(f"unsupported input_image_type {type(image)}")
            return input_image

97
98
99
100
101
102
103
104
105
106
107
108
109
110
        def convert_pt_to_type(image, input_image_type):
            if input_image_type == "pt":
                input_image = image
            elif input_image_type == "np":
                input_image = VaeImageProcessor.pt_to_numpy(image)
            elif input_image_type == "pil":
                input_image = VaeImageProcessor.pt_to_numpy(image)
                input_image = VaeImageProcessor.numpy_to_pil(input_image)
            else:
                raise ValueError(f"unsupported input_image_type {input_image_type}.")
            return input_image

        for image_param in self.image_params:
            if image_param in inputs.keys():
111
112
113
                inputs[image_param] = convert_pt_to_type(
                    convert_to_pt(inputs[image_param]).to(device), input_image_type
                )
114
115
116
117
118

        inputs["output_type"] = output_type

        return inputs

119
    def test_pt_np_pil_outputs_equivalent(self, expected_max_diff=1e-4):
120
121
122
        self._test_pt_np_pil_outputs_equivalent(expected_max_diff=expected_max_diff)

    def _test_pt_np_pil_outputs_equivalent(self, expected_max_diff=1e-4, input_image_type="pt"):
123
124
125
126
127
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

128
129
130
131
132
133
134
135
136
        output_pt = pipe(
            **self.get_dummy_inputs_by_type(torch_device, input_image_type=input_image_type, output_type="pt")
        )[0]
        output_np = pipe(
            **self.get_dummy_inputs_by_type(torch_device, input_image_type=input_image_type, output_type="np")
        )[0]
        output_pil = pipe(
            **self.get_dummy_inputs_by_type(torch_device, input_image_type=input_image_type, output_type="pil")
        )[0]
137
138

        max_diff = np.abs(output_pt.cpu().numpy().transpose(0, 2, 3, 1) - output_np).max()
139
140
141
        self.assertLess(
            max_diff, expected_max_diff, "`output_type=='pt'` generate different results from `output_type=='np'`"
        )
142
143

        max_diff = np.abs(np.array(output_pil[0]) - (output_np * 255).round()).max()
144
        self.assertLess(max_diff, 2.0, "`output_type=='pil'` generate different results from `output_type=='np'`")
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

    def test_pt_np_pil_inputs_equivalent(self):
        if len(self.image_params) == 0:
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        out_input_pt = pipe(**self.get_dummy_inputs_by_type(torch_device, input_image_type="pt"))[0]
        out_input_np = pipe(**self.get_dummy_inputs_by_type(torch_device, input_image_type="np"))[0]
        out_input_pil = pipe(**self.get_dummy_inputs_by_type(torch_device, input_image_type="pil"))[0]

        max_diff = np.abs(out_input_pt - out_input_np).max()
        self.assertLess(max_diff, 1e-4, "`input_type=='pt'` generate different result from `input_type=='np'`")
        max_diff = np.abs(out_input_pil - out_input_np).max()
        self.assertLess(max_diff, 1e-2, "`input_type=='pt'` generate different result from `input_type=='np'`")

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
    def test_latents_input(self):
        if len(self.image_latents_params) == 0:
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.image_processor = VaeImageProcessor(do_resize=False, do_normalize=False)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        out = pipe(**self.get_dummy_inputs_by_type(torch_device, input_image_type="pt"))[0]

        vae = components["vae"]
        inputs = self.get_dummy_inputs_by_type(torch_device, input_image_type="pt")
        generator = inputs["generator"]
        for image_param in self.image_latents_params:
            if image_param in inputs.keys():
                inputs[image_param] = (
                    vae.encode(inputs[image_param]).latent_dist.sample(generator) * vae.config.scaling_factor
                )
        out_latents_inputs = pipe(**inputs)[0]

        max_diff = np.abs(out - out_latents_inputs).max()
        self.assertLess(max_diff, 1e-4, "passing latents as image input generate different result from passing image")

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
    def test_multi_vae(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        block_out_channels = pipe.vae.config.block_out_channels
        norm_num_groups = pipe.vae.config.norm_num_groups

        vae_classes = [AutoencoderKL, AsymmetricAutoencoderKL, ConsistencyDecoderVAE, AutoencoderTiny]
        configs = [
            get_autoencoder_kl_config(block_out_channels, norm_num_groups),
            get_asym_autoencoder_kl_config(block_out_channels, norm_num_groups),
            get_consistency_vae_config(block_out_channels, norm_num_groups),
            get_autoencoder_tiny_config(block_out_channels),
        ]

        out_np = pipe(**self.get_dummy_inputs_by_type(torch_device, input_image_type="np"))[0]

        for vae_cls, config in zip(vae_classes, configs):
            vae = vae_cls(**config)
            vae = vae.to(torch_device)
            components["vae"] = vae
            vae_pipe = self.pipeline_class(**components)
            out_vae_np = vae_pipe(**self.get_dummy_inputs_by_type(torch_device, input_image_type="np"))[0]

            assert out_vae_np.shape == out_np.shape

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
@require_torch
class PipelineKarrasSchedulerTesterMixin:
    """
    This mixin is designed to be used with unittest.TestCase classes.
    It provides a set of common tests for each PyTorch pipeline that makes use of KarrasDiffusionSchedulers
    equivalence of dict and tuple outputs, etc.
    """

    def test_karras_schedulers_shape(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)

        # make sure that PNDM does not need warm-up
        pipe.scheduler.register_to_config(skip_prk_steps=True)

        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        inputs = self.get_dummy_inputs(torch_device)
        inputs["num_inference_steps"] = 2

        if "strength" in inputs:
            inputs["num_inference_steps"] = 4
            inputs["strength"] = 0.5

        outputs = []
        for scheduler_enum in KarrasDiffusionSchedulers:
            if "KDPM2" in scheduler_enum.name:
                inputs["num_inference_steps"] = 5

            scheduler_cls = getattr(diffusers, scheduler_enum.name)
            pipe.scheduler = scheduler_cls.from_config(pipe.scheduler.config)
            output = pipe(**inputs)[0]
            outputs.append(output)

            if "KDPM2" in scheduler_enum.name:
                inputs["num_inference_steps"] = 2

        assert check_same_shape(outputs)


258
259
260
261
262
263
264
265
@require_torch
class PipelineTesterMixin:
    """
    This mixin is designed to be used with unittest.TestCase classes.
    It provides a set of common tests for each PyTorch pipeline, e.g. saving and loading the pipeline,
    equivalence of dict and tuple outputs, etc.
    """

266
267
268
269
270
271
272
273
274
275
276
277
278
    # Canonical parameters that are passed to `__call__` regardless
    # of the type of pipeline. They are always optional and have common
    # sense default values.
    required_optional_params = frozenset(
        [
            "num_inference_steps",
            "num_images_per_prompt",
            "generator",
            "latents",
            "output_type",
            "return_dict",
        ]
    )
279

280
281
    # set these parameters to False in the child class if the pipeline does not support the corresponding functionality
    test_attention_slicing = True
282

283
284
    test_xformers_attention = True

285
286
287
288
289
    def get_generator(self, seed):
        device = torch_device if torch_device != "mps" else "cpu"
        generator = torch.Generator(device).manual_seed(seed)
        return generator

290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
    @property
    def pipeline_class(self) -> Union[Callable, DiffusionPipeline]:
        raise NotImplementedError(
            "You need to set the attribute `pipeline_class = ClassNameOfPipeline` in the child test class. "
            "See existing pipeline tests for reference."
        )

    def get_dummy_components(self):
        raise NotImplementedError(
            "You need to implement `get_dummy_components(self)` in the child test class. "
            "See existing pipeline tests for reference."
        )

    def get_dummy_inputs(self, device, seed=0):
        raise NotImplementedError(
            "You need to implement `get_dummy_inputs(self, device, seed)` in the child test class. "
            "See existing pipeline tests for reference."
        )

309
310
311
312
313
    @property
    def params(self) -> frozenset:
        raise NotImplementedError(
            "You need to set the attribute `params` in the child test class. "
            "`params` are checked for if all values are present in `__call__`'s signature."
314
            " You can set `params` using one of the common set of parameters defined in `pipeline_params.py`"
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
            " e.g., `TEXT_TO_IMAGE_PARAMS` defines the common parameters used in text to  "
            "image pipelines, including prompts and prompt embedding overrides."
            "If your pipeline's set of arguments has minor changes from one of the common sets of arguments, "
            "do not make modifications to the existing common sets of arguments. I.e. a text to image pipeline "
            "with non-configurable height and width arguments should set the attribute as "
            "`params = TEXT_TO_IMAGE_PARAMS - {'height', 'width'}`. "
            "See existing pipeline tests for reference."
        )

    @property
    def batch_params(self) -> frozenset:
        raise NotImplementedError(
            "You need to set the attribute `batch_params` in the child test class. "
            "`batch_params` are the parameters required to be batched when passed to the pipeline's "
            "`__call__` method. `pipeline_params.py` provides some common sets of parameters such as "
            "`TEXT_TO_IMAGE_BATCH_PARAMS`, `IMAGE_VARIATION_BATCH_PARAMS`, etc... If your pipeline's "
            "set of batch arguments has minor changes from one of the common sets of batch arguments, "
            "do not make modifications to the existing common sets of batch arguments. I.e. a text to "
            "image pipeline `negative_prompt` is not batched should set the attribute as "
            "`batch_params = TEXT_TO_IMAGE_BATCH_PARAMS - {'negative_prompt'}`. "
            "See existing pipeline tests for reference."
        )

338
339
340
341
342
343
344
345
346
347
348
349
350
351
    @property
    def callback_cfg_params(self) -> frozenset:
        raise NotImplementedError(
            "You need to set the attribute `callback_cfg_params` in the child test class that requires to run test_callback_cfg. "
            "`callback_cfg_params` are the parameters that needs to be passed to the pipeline's callback "
            "function when dynamically adjusting `guidance_scale`. They are variables that require special"
            "treatment when `do_classifier_free_guidance` is `True`. `pipeline_params.py` provides some common"
            " sets of parameters such as `TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS`. If your pipeline's "
            "set of cfg arguments has minor changes from one of the common sets of cfg arguments, "
            "do not make modifications to the existing common sets of cfg arguments. I.e. for inpaint pipeine, you "
            " need to adjust batch size of `mask` and `masked_image_latents` so should set the attribute as"
            "`callback_cfg_params = TEXT_TO_IMAGE_CFG_PARAMS.union({'mask', 'masked_image_latents'})`"
        )

352
353
354
355
356
357
    def tearDown(self):
        # clean up the VRAM after each test in case of CUDA runtime errors
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

358
    def test_save_load_local(self, expected_max_difference=5e-4):
359
360
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
361
362
363
364
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()

365
366
367
368
369
370
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        output = pipe(**inputs)[0]

371
372
373
        logger = logging.get_logger("diffusers.pipelines.pipeline_utils")
        logger.setLevel(diffusers.logging.INFO)

374
        with tempfile.TemporaryDirectory() as tmpdir:
375
            pipe.save_pretrained(tmpdir, safe_serialization=False)
376
377
378
379
380
381
382
383

            with CaptureLogger(logger) as cap_logger:
                pipe_loaded = self.pipeline_class.from_pretrained(tmpdir)

            for name in pipe_loaded.components.keys():
                if name not in pipe_loaded._optional_components:
                    assert name in str(cap_logger)

384
385
386
387
388
389
            pipe_loaded.to(torch_device)
            pipe_loaded.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        output_loaded = pipe_loaded(**inputs)[0]

390
        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
391
        self.assertLess(max_diff, expected_max_difference)
392

393
394
395
396
397
    def test_pipeline_call_signature(self):
        self.assertTrue(
            hasattr(self.pipeline_class, "__call__"), f"{self.pipeline_class} should have a `__call__` method"
        )

398
399
        parameters = inspect.signature(self.pipeline_class.__call__).parameters

400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
        optional_parameters = set()

        for k, v in parameters.items():
            if v.default != inspect._empty:
                optional_parameters.add(k)

        parameters = set(parameters.keys())
        parameters.remove("self")
        parameters.discard("kwargs")  # kwargs can be added if arguments of pipeline call function are deprecated

        remaining_required_parameters = set()

        for param in self.params:
            if param not in parameters:
                remaining_required_parameters.add(param)
415

416
417
418
419
420
421
        self.assertTrue(
            len(remaining_required_parameters) == 0,
            f"Required parameters not present: {remaining_required_parameters}",
        )

        remaining_required_optional_parameters = set()
422

423
        for param in self.required_optional_params:
424
425
426
427
428
429
430
            if param not in optional_parameters:
                remaining_required_optional_parameters.add(param)

        self.assertTrue(
            len(remaining_required_optional_parameters) == 0,
            f"Required optional parameters not present: {remaining_required_optional_parameters}",
        )
431

432
    def test_inference_batch_consistent(self, batch_sizes=[2]):
433
        self._test_inference_batch_consistent(batch_sizes=batch_sizes)
434

435
    def _test_inference_batch_consistent(
436
        self, batch_sizes=[2], additional_params_copy_to_batched_inputs=["num_inference_steps"]
437
    ):
438
439
440
441
442
443
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
444
        inputs["generator"] = self.get_generator(0)
445
446
447
448

        logger = logging.get_logger(pipe.__module__)
        logger.setLevel(level=diffusers.logging.FATAL)

449
450
        # prepare batched inputs
        batched_inputs = []
451
        for batch_size in batch_sizes:
452
453
            batched_input = {}
            batched_input.update(inputs)
454

455
456
457
            for name in self.batch_params:
                if name not in inputs:
                    continue
458

459
460
461
462
463
                value = inputs[name]
                if name == "prompt":
                    len_prompt = len(value)
                    # make unequal batch sizes
                    batched_input[name] = [value[: len_prompt // i] for i in range(1, batch_size + 1)]
464

465
466
                    # make last batch super long
                    batched_input[name][-1] = 100 * "very long"
467

468
469
                else:
                    batched_input[name] = batch_size * [value]
470

471
472
            if "generator" in inputs:
                batched_input["generator"] = [self.get_generator(i) for i in range(batch_size)]
473

474
475
            if "batch_size" in inputs:
                batched_input["batch_size"] = batch_size
476

477
            batched_inputs.append(batched_input)
478
479

        logger.setLevel(level=diffusers.logging.WARNING)
480
481
482
        for batch_size, batched_input in zip(batch_sizes, batched_inputs):
            output = pipe(**batched_input)
            assert len(output[0]) == batch_size
483

484
485
    def test_inference_batch_single_identical(self, batch_size=3, expected_max_diff=1e-4):
        self._test_inference_batch_single_identical(batch_size=batch_size, expected_max_diff=expected_max_diff)
486
487

    def _test_inference_batch_single_identical(
488
        self,
489
        batch_size=2,
490
        expected_max_diff=1e-4,
491
        additional_params_copy_to_batched_inputs=["num_inference_steps"],
492
    ):
493
494
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
495
496
497
498
        for components in pipe.components.values():
            if hasattr(components, "set_default_attn_processor"):
                components.set_default_attn_processor()

499
500
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
501
502
503
        inputs = self.get_dummy_inputs(torch_device)
        # Reset generator in case it is has been used in self.get_dummy_inputs
        inputs["generator"] = self.get_generator(0)
504
505
506
507
508
509

        logger = logging.get_logger(pipe.__module__)
        logger.setLevel(level=diffusers.logging.FATAL)

        # batchify inputs
        batched_inputs = {}
510
        batched_inputs.update(inputs)
511

512
513
514
        for name in self.batch_params:
            if name not in inputs:
                continue
515

516
517
518
519
520
            value = inputs[name]
            if name == "prompt":
                len_prompt = len(value)
                batched_inputs[name] = [value[: len_prompt // i] for i in range(1, batch_size + 1)]
                batched_inputs[name][-1] = 100 * "very long"
521

522
523
            else:
                batched_inputs[name] = batch_size * [value]
524

525
526
        if "generator" in inputs:
            batched_inputs["generator"] = [self.get_generator(i) for i in range(batch_size)]
527

528
529
530
531
532
        if "batch_size" in inputs:
            batched_inputs["batch_size"] = batch_size

        for arg in additional_params_copy_to_batched_inputs:
            batched_inputs[arg] = inputs[arg]
533
534

        output = pipe(**inputs)
535
        output_batch = pipe(**batched_inputs)
536

537
        assert output_batch[0].shape[0] == batch_size
538

YiYi Xu's avatar
YiYi Xu committed
539
        max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
540
        assert max_diff < expected_max_diff
541

542
    def test_dict_tuple_outputs_equivalent(self, expected_max_difference=1e-4):
543
544
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
545
546
547
548
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()

549
550
551
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

552
553
554
        generator_device = "cpu"
        output = pipe(**self.get_dummy_inputs(generator_device))[0]
        output_tuple = pipe(**self.get_dummy_inputs(generator_device), return_dict=False)[0]
555

556
        max_diff = np.abs(to_np(output) - to_np(output_tuple)).max()
557
        self.assertLess(max_diff, expected_max_difference)
558
559
560

    def test_components_function(self):
        init_components = self.get_dummy_components()
Kashif Rasul's avatar
Kashif Rasul committed
561
562
        init_components = {k: v for k, v in init_components.items() if not isinstance(v, (str, int, float))}

563
564
565
566
567
568
        pipe = self.pipeline_class(**init_components)

        self.assertTrue(hasattr(pipe, "components"))
        self.assertTrue(set(pipe.components.keys()) == set(init_components.keys()))

    @unittest.skipIf(torch_device != "cuda", reason="float16 requires CUDA")
569
    def test_float16_inference(self, expected_max_diff=5e-2):
570
571
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
572
573
574
575
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()

576
577
578
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

YiYi Xu's avatar
YiYi Xu committed
579
        components = self.get_dummy_components()
580
        pipe_fp16 = self.pipeline_class(**components)
581
582
583
584
        for component in pipe_fp16.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()

585
        pipe_fp16.to(torch_device, torch.float16)
586
587
        pipe_fp16.set_progress_bar_config(disable=None)

588
589
590
591
592
593
594
595
596
597
598
599
600
        inputs = self.get_dummy_inputs(torch_device)
        # Reset generator in case it is used inside dummy inputs
        if "generator" in inputs:
            inputs["generator"] = self.get_generator(0)

        output = pipe(**inputs)[0]

        fp16_inputs = self.get_dummy_inputs(torch_device)
        # Reset generator in case it is used inside dummy inputs
        if "generator" in fp16_inputs:
            fp16_inputs["generator"] = self.get_generator(0)

        output_fp16 = pipe_fp16(**fp16_inputs)[0]
601

602
        max_diff = np.abs(to_np(output) - to_np(output_fp16)).max()
603
        self.assertLess(max_diff, expected_max_diff, "The outputs of the fp16 and fp32 pipelines are too different.")
604
605

    @unittest.skipIf(torch_device != "cuda", reason="float16 requires CUDA")
606
    def test_save_load_float16(self, expected_max_diff=1e-2):
607
608
609
610
        components = self.get_dummy_components()
        for name, module in components.items():
            if hasattr(module, "half"):
                components[name] = module.to(torch_device).half()
611

612
        pipe = self.pipeline_class(**components)
613
614
615
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()
616
617
618
619
620
621
622
623
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        output = pipe(**inputs)[0]

        with tempfile.TemporaryDirectory() as tmpdir:
            pipe.save_pretrained(tmpdir)
624
            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir, torch_dtype=torch.float16)
625
626
627
            for component in pipe_loaded.components.values():
                if hasattr(component, "set_default_attn_processor"):
                    component.set_default_attn_processor()
628
629
630
631
632
633
634
635
636
637
638
639
            pipe_loaded.to(torch_device)
            pipe_loaded.set_progress_bar_config(disable=None)

        for name, component in pipe_loaded.components.items():
            if hasattr(component, "dtype"):
                self.assertTrue(
                    component.dtype == torch.float16,
                    f"`{name}.dtype` switched from `float16` to {component.dtype} after loading.",
                )

        inputs = self.get_dummy_inputs(torch_device)
        output_loaded = pipe_loaded(**inputs)[0]
640
        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
641
642
643
        self.assertLess(
            max_diff, expected_max_diff, "The output of the fp16 pipeline changed after saving and loading."
        )
644

645
    def test_save_load_optional_components(self, expected_max_difference=1e-4):
646
647
648
649
650
        if not hasattr(self.pipeline_class, "_optional_components"):
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
651
652
653
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()
654
655
656
657
658
659
660
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        # set all optional components to None
        for optional_component in pipe._optional_components:
            setattr(pipe, optional_component, None)

Dhruv Nair's avatar
Dhruv Nair committed
661
662
        generator_device = "cpu"
        inputs = self.get_dummy_inputs(generator_device)
663
664
665
        output = pipe(**inputs)[0]

        with tempfile.TemporaryDirectory() as tmpdir:
666
            pipe.save_pretrained(tmpdir, safe_serialization=False)
667
            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir)
668
669
670
            for component in pipe_loaded.components.values():
                if hasattr(component, "set_default_attn_processor"):
                    component.set_default_attn_processor()
671
672
673
674
675
676
677
678
679
            pipe_loaded.to(torch_device)
            pipe_loaded.set_progress_bar_config(disable=None)

        for optional_component in pipe._optional_components:
            self.assertTrue(
                getattr(pipe_loaded, optional_component) is None,
                f"`{optional_component}` did not stay set to None after loading.",
            )

Dhruv Nair's avatar
Dhruv Nair committed
680
        inputs = self.get_dummy_inputs(generator_device)
681
682
        output_loaded = pipe_loaded(**inputs)[0]

683
        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
684
        self.assertLess(max_diff, expected_max_difference)
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703

    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
    def test_to_device(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.set_progress_bar_config(disable=None)

        pipe.to("cpu")
        model_devices = [component.device.type for component in components.values() if hasattr(component, "device")]
        self.assertTrue(all(device == "cpu" for device in model_devices))

        output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
        self.assertTrue(np.isnan(output_cpu).sum() == 0)

        pipe.to("cuda")
        model_devices = [component.device.type for component in components.values() if hasattr(component, "device")]
        self.assertTrue(all(device == "cuda" for device in model_devices))

        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
704
        self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
705

706
707
708
709
710
711
712
713
714
715
716
717
    def test_to_dtype(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.set_progress_bar_config(disable=None)

        model_dtypes = [component.dtype for component in components.values() if hasattr(component, "dtype")]
        self.assertTrue(all(dtype == torch.float32 for dtype in model_dtypes))

        pipe.to(torch_dtype=torch.float16)
        model_dtypes = [component.dtype for component in components.values() if hasattr(component, "dtype")]
        self.assertTrue(all(dtype == torch.float16 for dtype in model_dtypes))

718
719
    def test_attention_slicing_forward_pass(self, expected_max_diff=1e-3):
        self._test_attention_slicing_forward_pass(expected_max_diff=expected_max_diff)
720

721
722
723
    def _test_attention_slicing_forward_pass(
        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
    ):
724
725
726
727
728
        if not self.test_attention_slicing:
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
729
730
731
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()
732
733
734
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

735
736
        generator_device = "cpu"
        inputs = self.get_dummy_inputs(generator_device)
737
738
739
        output_without_slicing = pipe(**inputs)[0]

        pipe.enable_attention_slicing(slice_size=1)
740
        inputs = self.get_dummy_inputs(generator_device)
741
742
        output_with_slicing = pipe(**inputs)[0]

743
        if test_max_difference:
744
            max_diff = np.abs(to_np(output_with_slicing) - to_np(output_without_slicing)).max()
745
            self.assertLess(max_diff, expected_max_diff, "Attention slicing should not affect the inference results")
746

747
        if test_mean_pixel_difference:
YiYi Xu's avatar
YiYi Xu committed
748
            assert_mean_pixel_difference(to_np(output_with_slicing[0]), to_np(output_without_slicing[0]))
749
750

    @unittest.skipIf(
751
752
        torch_device != "cuda" or not is_accelerate_available() or is_accelerate_version("<", "0.14.0"),
        reason="CPU offload is only available with CUDA and `accelerate v0.14.0` or higher",
753
    )
754
    def test_sequential_cpu_offload_forward_pass(self, expected_max_diff=1e-4):
755
756
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
757
758
759
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()
760
761
762
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

763
764
        generator_device = "cpu"
        inputs = self.get_dummy_inputs(generator_device)
765
766
767
        output_without_offload = pipe(**inputs)[0]

        pipe.enable_sequential_cpu_offload()
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795

        inputs = self.get_dummy_inputs(generator_device)
        output_with_offload = pipe(**inputs)[0]

        max_diff = np.abs(to_np(output_with_offload) - to_np(output_without_offload)).max()
        self.assertLess(max_diff, expected_max_diff, "CPU offloading should not affect the inference results")

    @unittest.skipIf(
        torch_device != "cuda" or not is_accelerate_available() or is_accelerate_version("<", "0.17.0"),
        reason="CPU offload is only available with CUDA and `accelerate v0.17.0` or higher",
    )
    def test_model_cpu_offload_forward_pass(self, expected_max_diff=2e-4):
        generator_device = "cpu"
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)

        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()

        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(generator_device)
        output_without_offload = pipe(**inputs)[0]

        pipe.enable_model_cpu_offload()
        inputs = self.get_dummy_inputs(generator_device)
796
797
        output_with_offload = pipe(**inputs)[0]

798
        max_diff = np.abs(to_np(output_with_offload) - to_np(output_without_offload)).max()
799
        self.assertLess(max_diff, expected_max_diff, "CPU offloading should not affect the inference results")
800
801
802
803
804
        offloaded_modules = [
            v
            for k, v in pipe.components.items()
            if isinstance(v, torch.nn.Module) and k not in pipe._exclude_from_cpu_offload
        ]
805
806
807
808
        (
            self.assertTrue(all(v.device.type == "cpu" for v in offloaded_modules)),
            f"Not offloaded: {[v for v in offloaded_modules if v.device.type != 'cpu']}",
        )
809
810
811
812
813

    @unittest.skipIf(
        torch_device != "cuda" or not is_xformers_available(),
        reason="XFormers attention is only available with CUDA and `xformers` installed",
    )
Kashif Rasul's avatar
Kashif Rasul committed
814
    def test_xformers_attention_forwardGenerator_pass(self):
Will Berman's avatar
Will Berman committed
815
816
        self._test_xformers_attention_forwardGenerator_pass()

817
818
819
    def _test_xformers_attention_forwardGenerator_pass(
        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-4
    ):
820
821
822
823
824
        if not self.test_xformers_attention:
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
825
826
827
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()
828
829
830
831
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
832
        output_without_offload = pipe(**inputs)[0]
833
834
835
        output_without_offload = (
            output_without_offload.cpu() if torch.is_tensor(output_without_offload) else output_without_offload
        )
836
837
838

        pipe.enable_xformers_memory_efficient_attention()
        inputs = self.get_dummy_inputs(torch_device)
839
        output_with_offload = pipe(**inputs)[0]
840
841
842
        output_with_offload = (
            output_with_offload.cpu() if torch.is_tensor(output_with_offload) else output_without_offload
        )
843

Will Berman's avatar
Will Berman committed
844
        if test_max_difference:
845
            max_diff = np.abs(to_np(output_with_offload) - to_np(output_without_offload)).max()
846
            self.assertLess(max_diff, expected_max_diff, "XFormers attention should not affect the inference results")
Will Berman's avatar
Will Berman committed
847

848
849
        if test_mean_pixel_difference:
            assert_mean_pixel_difference(output_with_offload[0], output_without_offload[0])
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871

    def test_progress_bar(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.to(torch_device)

        inputs = self.get_dummy_inputs(torch_device)
        with io.StringIO() as stderr, contextlib.redirect_stderr(stderr):
            _ = pipe(**inputs)
            stderr = stderr.getvalue()
            # we can't calculate the number of progress steps beforehand e.g. for strength-dependent img2img,
            # so we just match "5" in "#####| 1/5 [00:01<00:00]"
            max_steps = re.search("/(.*?) ", stderr).group(1)
            self.assertTrue(max_steps is not None and len(max_steps) > 0)
            self.assertTrue(
                f"{max_steps}/{max_steps}" in stderr, "Progress bar should be enabled and stopped at the max step"
            )

        pipe.set_progress_bar_config(disable=True)
        with io.StringIO() as stderr, contextlib.redirect_stderr(stderr):
            _ = pipe(**inputs)
            self.assertTrue(stderr.getvalue() == "", "Progress bar should be disabled")
872

873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
    def test_num_images_per_prompt(self):
        sig = inspect.signature(self.pipeline_class.__call__)

        if "num_images_per_prompt" not in sig.parameters:
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        batch_sizes = [1, 2]
        num_images_per_prompts = [1, 2]

        for batch_size in batch_sizes:
            for num_images_per_prompt in num_images_per_prompts:
                inputs = self.get_dummy_inputs(torch_device)

                for key in inputs.keys():
                    if key in self.batch_params:
                        inputs[key] = batch_size * [inputs[key]]

895
                images = pipe(**inputs, num_images_per_prompt=num_images_per_prompt)[0]
896
897
898

                assert images.shape[0] == batch_size * num_images_per_prompt

899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
    def test_cfg(self):
        sig = inspect.signature(self.pipeline_class.__call__)

        if "guidance_scale" not in sig.parameters:
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)

        inputs["guidance_scale"] = 1.0
        out_no_cfg = pipe(**inputs)[0]

        inputs["guidance_scale"] = 7.5
        out_cfg = pipe(**inputs)[0]

        assert out_cfg.shape == out_no_cfg.shape

920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
    def test_callback_inputs(self):
        sig = inspect.signature(self.pipeline_class.__call__)
        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
        has_callback_step_end = "callback_on_step_end" in sig.parameters

        if not (has_callback_tensor_inputs and has_callback_step_end):
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        self.assertTrue(
            hasattr(pipe, "_callback_tensor_inputs"),
            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
        )

        def callback_inputs_subset(pipe, i, t, callback_kwargs):
            # interate over callback args
            for tensor_name, tensor_value in callback_kwargs.items():
                # check that we're only passing in allowed tensor inputs
                assert tensor_name in pipe._callback_tensor_inputs

            return callback_kwargs

        def callback_inputs_all(pipe, i, t, callback_kwargs):
            for tensor_name in pipe._callback_tensor_inputs:
                assert tensor_name in callback_kwargs

            # interate over callback args
            for tensor_name, tensor_value in callback_kwargs.items():
                # check that we're only passing in allowed tensor inputs
                assert tensor_name in pipe._callback_tensor_inputs

            return callback_kwargs

        inputs = self.get_dummy_inputs(torch_device)

        # Test passing in a subset
        inputs["callback_on_step_end"] = callback_inputs_subset
        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
        inputs["output_type"] = "latent"
        output = pipe(**inputs)[0]

        # Test passing in a everything
        inputs["callback_on_step_end"] = callback_inputs_all
        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
        inputs["output_type"] = "latent"
        output = pipe(**inputs)[0]

        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
            is_last = i == (pipe.num_timesteps - 1)
            if is_last:
                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
            return callback_kwargs

        inputs["callback_on_step_end"] = callback_inputs_change_tensor
        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
        inputs["output_type"] = "latent"
        output = pipe(**inputs)[0]
        assert output.abs().sum() == 0

    def test_callback_cfg(self):
        sig = inspect.signature(self.pipeline_class.__call__)
        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
        has_callback_step_end = "callback_on_step_end" in sig.parameters

        if not (has_callback_tensor_inputs and has_callback_step_end):
            return

        if "guidance_scale" not in sig.parameters:
            return

        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        self.assertTrue(
            hasattr(pipe, "_callback_tensor_inputs"),
            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
        )

        def callback_increase_guidance(pipe, i, t, callback_kwargs):
            pipe._guidance_scale += 1.0

            return callback_kwargs

        inputs = self.get_dummy_inputs(torch_device)

        # use cfg guidance because some pipelines modify the shape of the latents
        # outside of the denoising loop
        inputs["guidance_scale"] = 2.0
        inputs["callback_on_step_end"] = callback_increase_guidance
        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
        _ = pipe(**inputs)[0]

        # we increase the guidance scale by 1.0 at every step
        # check that the guidance scale is increased by the number of scheduler timesteps
        # accounts for models that modify the number of inference steps based on strength
        assert pipe.guidance_scale == (inputs["guidance_scale"] + pipe.num_timesteps)

1021

1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
@is_staging_test
class PipelinePushToHubTester(unittest.TestCase):
    identifier = uuid.uuid4()
    repo_id = f"test-pipeline-{identifier}"
    org_repo_id = f"valid_org/{repo_id}-org"

    def get_pipeline_components(self):
        unet = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )

        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="scaled_linear",
            clip_sample=False,
            set_alpha_to_one=False,
        )

        vae = AutoencoderKL(
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=4,
        )

        text_encoder_config = CLIPTextConfig(
            bos_token_id=0,
            eos_token_id=2,
            hidden_size=32,
            intermediate_size=37,
            layer_norm_eps=1e-05,
            num_attention_heads=4,
            num_hidden_layers=5,
            pad_token_id=1,
            vocab_size=1000,
        )
        text_encoder = CLIPTextModel(text_encoder_config)

        with tempfile.TemporaryDirectory() as tmpdir:
            dummy_vocab = {"<|startoftext|>": 0, "<|endoftext|>": 1, "!": 2}
            vocab_path = os.path.join(tmpdir, "vocab.json")
            with open(vocab_path, "w") as f:
                json.dump(dummy_vocab, f)

            merges = "Ġ t\nĠt h"
            merges_path = os.path.join(tmpdir, "merges.txt")
            with open(merges_path, "w") as f:
                f.writelines(merges)
            tokenizer = CLIPTokenizer(vocab_file=vocab_path, merges_file=merges_path)

        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
            "feature_extractor": None,
        }
        return components

    def test_push_to_hub(self):
        components = self.get_pipeline_components()
        pipeline = StableDiffusionPipeline(**components)
        pipeline.push_to_hub(self.repo_id, token=TOKEN)

        new_model = UNet2DConditionModel.from_pretrained(f"{USER}/{self.repo_id}", subfolder="unet")
        unet = components["unet"]
        for p1, p2 in zip(unet.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(token=TOKEN, repo_id=self.repo_id)

        # Push to hub via save_pretrained
        with tempfile.TemporaryDirectory() as tmp_dir:
            pipeline.save_pretrained(tmp_dir, repo_id=self.repo_id, push_to_hub=True, token=TOKEN)

        new_model = UNet2DConditionModel.from_pretrained(f"{USER}/{self.repo_id}", subfolder="unet")
        for p1, p2 in zip(unet.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(self.repo_id, token=TOKEN)

    def test_push_to_hub_in_organization(self):
        components = self.get_pipeline_components()
        pipeline = StableDiffusionPipeline(**components)
        pipeline.push_to_hub(self.org_repo_id, token=TOKEN)

        new_model = UNet2DConditionModel.from_pretrained(self.org_repo_id, subfolder="unet")
        unet = components["unet"]
        for p1, p2 in zip(unet.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(token=TOKEN, repo_id=self.org_repo_id)

        # Push to hub via save_pretrained
        with tempfile.TemporaryDirectory() as tmp_dir:
            pipeline.save_pretrained(tmp_dir, push_to_hub=True, token=TOKEN, repo_id=self.org_repo_id)

        new_model = UNet2DConditionModel.from_pretrained(self.org_repo_id, subfolder="unet")
        for p1, p2 in zip(unet.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(self.org_repo_id, token=TOKEN)


1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
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
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
# For SDXL and its derivative pipelines (such as ControlNet), we have the text encoders
# and the tokenizers as optional components. So, we need to override the `test_save_load_optional_components()`
# test for all such pipelines. This requires us to use a custom `encode_prompt()` function.
class SDXLOptionalComponentsTesterMixin:
    def encode_prompt(
        self, tokenizers, text_encoders, prompt: str, num_images_per_prompt: int = 1, negative_prompt: str = None
    ):
        device = text_encoders[0].device

        if isinstance(prompt, str):
            prompt = [prompt]
        batch_size = len(prompt)

        prompt_embeds_list = []
        for tokenizer, text_encoder in zip(tokenizers, text_encoders):
            text_inputs = tokenizer(
                prompt,
                padding="max_length",
                max_length=tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )

            text_input_ids = text_inputs.input_ids

            prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
            pooled_prompt_embeds = prompt_embeds[0]
            prompt_embeds = prompt_embeds.hidden_states[-2]
            prompt_embeds_list.append(prompt_embeds)

        prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)

        if negative_prompt is None:
            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
        else:
            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt

            negative_prompt_embeds_list = []
            for tokenizer, text_encoder in zip(tokenizers, text_encoders):
                uncond_input = tokenizer(
                    negative_prompt,
                    padding="max_length",
                    max_length=tokenizer.model_max_length,
                    truncation=True,
                    return_tensors="pt",
                )

                negative_prompt_embeds = text_encoder(uncond_input.input_ids.to(device), output_hidden_states=True)
                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
                negative_prompt_embeds_list.append(negative_prompt_embeds)

            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)

        bs_embed, seq_len, _ = prompt_embeds.shape

        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)

        # for classifier-free guidance
        # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
        seq_len = negative_prompt_embeds.shape[1]

        negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
        negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
            bs_embed * num_images_per_prompt, -1
        )

        # for classifier-free guidance
        negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
            bs_embed * num_images_per_prompt, -1
        )

        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds

    def _test_save_load_optional_components(self, expected_max_difference=1e-4):
        components = self.get_dummy_components()

        pipe = self.pipeline_class(**components)
        for optional_component in pipe._optional_components:
            setattr(pipe, optional_component, None)

        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        generator_device = "cpu"
        inputs = self.get_dummy_inputs(generator_device)

        tokenizer = components.pop("tokenizer")
        tokenizer_2 = components.pop("tokenizer_2")
        text_encoder = components.pop("text_encoder")
        text_encoder_2 = components.pop("text_encoder_2")

        tokenizers = [tokenizer, tokenizer_2] if tokenizer is not None else [tokenizer_2]
        text_encoders = [text_encoder, text_encoder_2] if text_encoder is not None else [text_encoder_2]
        prompt = inputs.pop("prompt")
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = self.encode_prompt(tokenizers, text_encoders, prompt)
        inputs["prompt_embeds"] = prompt_embeds
        inputs["negative_prompt_embeds"] = negative_prompt_embeds
        inputs["pooled_prompt_embeds"] = pooled_prompt_embeds
        inputs["negative_pooled_prompt_embeds"] = negative_pooled_prompt_embeds

        output = pipe(**inputs)[0]

        with tempfile.TemporaryDirectory() as tmpdir:
            pipe.save_pretrained(tmpdir)
            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir)
            for component in pipe_loaded.components.values():
                if hasattr(component, "set_default_attn_processor"):
                    component.set_default_attn_processor()
            pipe_loaded.to(torch_device)
            pipe_loaded.set_progress_bar_config(disable=None)

        for optional_component in pipe._optional_components:
            self.assertTrue(
                getattr(pipe_loaded, optional_component) is None,
                f"`{optional_component}` did not stay set to None after loading.",
            )

        inputs = self.get_dummy_inputs(generator_device)
        _ = inputs.pop("prompt")
        inputs["prompt_embeds"] = prompt_embeds
        inputs["negative_prompt_embeds"] = negative_prompt_embeds
        inputs["pooled_prompt_embeds"] = pooled_prompt_embeds
        inputs["negative_pooled_prompt_embeds"] = negative_pooled_prompt_embeds

        output_loaded = pipe_loaded(**inputs)[0]

        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
        self.assertLess(max_diff, expected_max_difference)


1286
1287
1288
# Some models (e.g. unCLIP) are extremely likely to significantly deviate depending on which hardware is used.
# This helper function is used to check that the image doesn't deviate on average more than 10 pixels from a
# reference image.
1289
def assert_mean_pixel_difference(image, expected_image, expected_max_diff=10):
1290
1291
1292
    image = np.asarray(DiffusionPipeline.numpy_to_pil(image)[0], dtype=np.float32)
    expected_image = np.asarray(DiffusionPipeline.numpy_to_pil(expected_image)[0], dtype=np.float32)
    avg_diff = np.abs(image - expected_image).mean()
1293
    assert avg_diff < expected_max_diff, f"Error image deviates {avg_diff} pixels on average"