pipeline_utils.py 100 KB
Newer Older
1
# coding=utf-8
Patrick von Platen's avatar
Patrick von Platen committed
2
# Copyright 2023 The HuggingFace Inc. team.
3
4
5
6
7
8
9
10
11
12
13
14
15
# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
#
# 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
import fnmatch
17
18
19
import importlib
import inspect
import os
20
import re
21
import sys
22
import warnings
23
24
from dataclasses import dataclass
from pathlib import Path
25
from typing import Any, Callable, Dict, List, Optional, Union
26
27

import numpy as np
Anh71me's avatar
Anh71me committed
28
import PIL.Image
29
import requests
30
import torch
31
32
33
34
35
36
37
from huggingface_hub import (
    ModelCard,
    create_repo,
    hf_hub_download,
    model_info,
    snapshot_download,
)
38
from huggingface_hub.utils import OfflineModeIsEnabled, validate_hf_hub_args
39
from packaging import version
40
from requests.exceptions import HTTPError
41
42
from tqdm.auto import tqdm

43
from .. import __version__
44
45
46
47
48
from ..configuration_utils import ConfigMixin
from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
from ..utils import (
    CONFIG_NAME,
49
50
    DEPRECATED_REVISION_ARGS,
    SAFETENSORS_WEIGHTS_NAME,
51
52
53
54
55
    WEIGHTS_NAME,
    BaseOutput,
    deprecate,
    get_class_from_dynamic_module,
    is_accelerate_available,
56
    is_accelerate_version,
57
    is_peft_available,
58
59
60
    is_torch_version,
    is_transformers_available,
    logging,
Patrick von Platen's avatar
Patrick von Platen committed
61
    numpy_to_pil,
62
)
Dhruv Nair's avatar
Dhruv Nair committed
63
from ..utils.torch_utils import is_compiled_module
64
65
66
67
68


if is_transformers_available():
    import transformers
    from transformers import PreTrainedModel
69
70
71
72
    from transformers.utils import FLAX_WEIGHTS_NAME as TRANSFORMERS_FLAX_WEIGHTS_NAME
    from transformers.utils import SAFE_WEIGHTS_NAME as TRANSFORMERS_SAFE_WEIGHTS_NAME
    from transformers.utils import WEIGHTS_NAME as TRANSFORMERS_WEIGHTS_NAME

73
from ..utils import FLAX_WEIGHTS_NAME, ONNX_EXTERNAL_WEIGHTS_NAME, ONNX_WEIGHTS_NAME, PushToHubMixin
74
75


76
77
78
79
if is_accelerate_available():
    import accelerate


80
81
82
83
INDEX_FILE = "diffusion_pytorch_model.bin"
CUSTOM_PIPELINE_FILE_NAME = "pipeline.py"
DUMMY_MODULES_FOLDER = "diffusers.utils"
TRANSFORMERS_DUMMY_MODULES_FOLDER = "transformers.utils"
84
CONNECTED_PIPES_KEYS = ["prior"]
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121


logger = logging.get_logger(__name__)


LOADABLE_CLASSES = {
    "diffusers": {
        "ModelMixin": ["save_pretrained", "from_pretrained"],
        "SchedulerMixin": ["save_pretrained", "from_pretrained"],
        "DiffusionPipeline": ["save_pretrained", "from_pretrained"],
        "OnnxRuntimeModel": ["save_pretrained", "from_pretrained"],
    },
    "transformers": {
        "PreTrainedTokenizer": ["save_pretrained", "from_pretrained"],
        "PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"],
        "PreTrainedModel": ["save_pretrained", "from_pretrained"],
        "FeatureExtractionMixin": ["save_pretrained", "from_pretrained"],
        "ProcessorMixin": ["save_pretrained", "from_pretrained"],
        "ImageProcessingMixin": ["save_pretrained", "from_pretrained"],
    },
    "onnxruntime.training": {
        "ORTModule": ["save_pretrained", "from_pretrained"],
    },
}

ALL_IMPORTABLE_CLASSES = {}
for library in LOADABLE_CLASSES:
    ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])


@dataclass
class ImagePipelineOutput(BaseOutput):
    """
    Output class for image pipelines.

    Args:
        images (`List[PIL.Image.Image]` or `np.ndarray`)
Steven Liu's avatar
Steven Liu committed
122
123
            List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
            num_channels)`.
124
125
126
127
128
129
130
131
132
133
134
135
    """

    images: Union[List[PIL.Image.Image], np.ndarray]


@dataclass
class AudioPipelineOutput(BaseOutput):
    """
    Output class for audio pipelines.

    Args:
        audios (`np.ndarray`)
Steven Liu's avatar
Steven Liu committed
136
            List of denoised audio samples of a NumPy array of shape `(batch_size, num_channels, sample_rate)`.
137
138
139
140
141
    """

    audios: np.ndarray


142
def is_safetensors_compatible(filenames, variant=None, passed_components=None) -> bool:
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
    """
    Checking for safetensors compatibility:
    - By default, all models are saved with the default pytorch serialization, so we use the list of default pytorch
      files to know which safetensors files are needed.
    - The model is safetensors compatible only if there is a matching safetensors file for every default pytorch file.

    Converting default pytorch serialized filenames to safetensors serialized filenames:
    - For models from the diffusers library, just replace the ".bin" extension with ".safetensors"
    - For models from the transformers library, the filename changes from "pytorch_model" to "model", and the ".bin"
      extension is replaced with ".safetensors"
    """
    pt_filenames = []

    sf_filenames = set()

158
159
    passed_components = passed_components or []

160
161
162
    for filename in filenames:
        _, extension = os.path.splitext(filename)

163
164
165
        if len(filename.split("/")) == 2 and filename.split("/")[0] in passed_components:
            continue

166
        if extension == ".bin":
167
            pt_filenames.append(os.path.normpath(filename))
168
        elif extension == ".safetensors":
169
            sf_filenames.add(os.path.normpath(filename))
170
171
172
173
174
175

    for filename in pt_filenames:
        #  filename = 'foo/bar/baz.bam' -> path = 'foo/bar', filename = 'baz', extention = '.bam'
        path, filename = os.path.split(filename)
        filename, extension = os.path.splitext(filename)

176
177
        if filename.startswith("pytorch_model"):
            filename = filename.replace("pytorch_model", "model")
178
        else:
179
180
            filename = filename

181
        expected_sf_filename = os.path.normpath(os.path.join(path, filename))
182
183
184
185
186
187
        expected_sf_filename = f"{expected_sf_filename}.safetensors"
        if expected_sf_filename not in sf_filenames:
            logger.warning(f"{expected_sf_filename} not found")
            return False

    return True
188
189


190
def variant_compatible_siblings(filenames, variant=None) -> Union[List[os.PathLike], str]:
191
192
193
194
195
196
197
    weight_names = [
        WEIGHTS_NAME,
        SAFETENSORS_WEIGHTS_NAME,
        FLAX_WEIGHTS_NAME,
        ONNX_WEIGHTS_NAME,
        ONNX_EXTERNAL_WEIGHTS_NAME,
    ]
198
199
200
201
202
203
204
205

    if is_transformers_available():
        weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME, TRANSFORMERS_FLAX_WEIGHTS_NAME]

    # model_pytorch, diffusion_model_pytorch, ...
    weight_prefixes = [w.split(".")[0] for w in weight_names]
    # .bin, .safetensors, ...
    weight_suffixs = [w.split(".")[-1] for w in weight_names]
206
    # -00001-of-00002
207
    transformers_index_format = r"\d{5}-of-\d{5}"
208
209

    if variant is not None:
210
        # `diffusion_pytorch_model.fp16.bin` as well as `model.fp16-00001-of-00002.safetensors`
211
        variant_file_re = re.compile(
212
            rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$"
213
214
215
        )
        # `text_encoder/pytorch_model.bin.index.fp16.json`
        variant_index_re = re.compile(
216
            rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.{variant}\.json$"
217
        )
218

219
    # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors`
220
    non_variant_file_re = re.compile(
221
        rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$"
222
    )
223
    # `text_encoder/pytorch_model.bin.index.json`
224
    non_variant_index_re = re.compile(rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.json")
225
226

    if variant is not None:
227
228
229
        variant_weights = {f for f in filenames if variant_file_re.match(f.split("/")[-1]) is not None}
        variant_indexes = {f for f in filenames if variant_index_re.match(f.split("/")[-1]) is not None}
        variant_filenames = variant_weights | variant_indexes
230
231
232
    else:
        variant_filenames = set()

233
234
235
    non_variant_weights = {f for f in filenames if non_variant_file_re.match(f.split("/")[-1]) is not None}
    non_variant_indexes = {f for f in filenames if non_variant_index_re.match(f.split("/")[-1]) is not None}
    non_variant_filenames = non_variant_weights | non_variant_indexes
236

237
    # all variant filenames will be used by default
238
    usable_filenames = set(variant_filenames)
239
240
241
242
243
244
245
246
247
248

    def convert_to_variant(filename):
        if "index" in filename:
            variant_filename = filename.replace("index", f"index.{variant}")
        elif re.compile(f"^(.*?){transformers_index_format}").match(filename) is not None:
            variant_filename = f"{filename.split('-')[0]}.{variant}-{'-'.join(filename.split('-')[1:])}"
        else:
            variant_filename = f"{filename.split('.')[0]}.{variant}.{filename.split('.')[1]}"
        return variant_filename

249
    for f in non_variant_filenames:
250
        variant_filename = convert_to_variant(f)
251
252
253
254
255
256
        if variant_filename not in usable_filenames:
            usable_filenames.add(f)

    return usable_filenames, variant_filenames


257
258
@validate_hf_hub_args
def warn_deprecated_model_variant(pretrained_model_name_or_path, token, variant, revision, model_filenames):
259
260
    info = model_info(
        pretrained_model_name_or_path,
261
        token=token,
262
263
        revision=None,
    )
264
    filenames = {sibling.rfilename for sibling in info.siblings}
265
266
267
    comp_model_filenames, _ = variant_compatible_siblings(filenames, variant=revision)
    comp_model_filenames = [".".join(f.split(".")[:1] + f.split(".")[2:]) for f in comp_model_filenames]

268
    if set(model_filenames).issubset(set(comp_model_filenames)):
269
270
271
272
273
274
275
276
277
278
279
        warnings.warn(
            f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'` even though you can load it via `variant=`{revision}`. Loading model variants via `revision='{revision}'` is deprecated and will be removed in diffusers v1. Please use `variant='{revision}'` instead.",
            FutureWarning,
        )
    else:
        warnings.warn(
            f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'`. This behavior is deprecated and will be removed in diffusers v1. One should use `variant='{revision}'` instead. However, it appears that {pretrained_model_name_or_path} currently does not have the required variant filenames in the 'main' branch. \n The Diffusers team and community would be very grateful if you could open an issue: https://github.com/huggingface/diffusers/issues/new with the title '{pretrained_model_name_or_path} is missing {revision} files' so that the correct variant file can be added.",
            FutureWarning,
        )


280
281
282
283
284
285
286
287
288
289
290
291
292
293
def _unwrap_model(model):
    """Unwraps a model."""
    if is_compiled_module(model):
        model = model._orig_mod

    if is_peft_available():
        from peft import PeftModel

        if isinstance(model, PeftModel):
            model = model.base_model.model

    return model


294
295
296
297
298
299
300
301
302
303
304
305
306
307
def maybe_raise_or_warn(
    library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module
):
    """Simple helper method to raise or warn in case incorrect module has been passed"""
    if not is_pipeline_module:
        library = importlib.import_module(library_name)
        class_obj = getattr(library, class_name)
        class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()}

        expected_class_obj = None
        for class_name, class_candidate in class_candidates.items():
            if class_candidate is not None and issubclass(class_obj, class_candidate):
                expected_class_obj = class_candidate

308
309
310
        # Dynamo wraps the original model in a private class.
        # I didn't find a public API to get the original class.
        sub_model = passed_class_obj[name]
311
312
        unwrapped_sub_model = _unwrap_model(sub_model)
        model_cls = unwrapped_sub_model.__class__
313
314

        if not issubclass(model_cls, expected_class_obj):
315
            raise ValueError(
316
                f"{passed_class_obj[name]} is of type: {model_cls}, but should be" f" {expected_class_obj}"
317
318
319
320
321
322
323
324
            )
    else:
        logger.warning(
            f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it"
            " has the correct type"
        )


325
326
327
def get_class_obj_and_candidates(
    library_name, class_name, importable_classes, pipelines, is_pipeline_module, component_name=None, cache_dir=None
):
328
    """Simple helper method to retrieve class object of module as well as potential parent class objects"""
329
330
    component_folder = os.path.join(cache_dir, component_name)

331
332
333
334
335
    if is_pipeline_module:
        pipeline_module = getattr(pipelines, library_name)

        class_obj = getattr(pipeline_module, class_name)
        class_candidates = {c: class_obj for c in importable_classes.keys()}
336
337
338
339
340
341
    elif os.path.isfile(os.path.join(component_folder, library_name + ".py")):
        # load custom component
        class_obj = get_class_from_dynamic_module(
            component_folder, module_file=library_name + ".py", class_name=class_name
        )
        class_candidates = {c: class_obj for c in importable_classes.keys()}
342
343
344
345
346
347
348
349
350
351
    else:
        # else we just import it from the library.
        library = importlib.import_module(library_name)

        class_obj = getattr(library, class_name)
        class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()}

    return class_obj, class_candidates


352
def _get_pipeline_class(
353
354
355
356
357
358
359
360
361
    class_obj,
    config,
    load_connected_pipeline=False,
    custom_pipeline=None,
    repo_id=None,
    hub_revision=None,
    class_name=None,
    cache_dir=None,
    revision=None,
362
):
363
364
365
366
367
368
    if custom_pipeline is not None:
        if custom_pipeline.endswith(".py"):
            path = Path(custom_pipeline)
            # decompose into folder & file
            file_name = path.name
            custom_pipeline = path.parent.absolute()
369
370
371
        elif repo_id is not None:
            file_name = f"{custom_pipeline}.py"
            custom_pipeline = repo_id
372
373
374
        else:
            file_name = CUSTOM_PIPELINE_FILE_NAME

375
376
377
378
379
        if repo_id is not None and hub_revision is not None:
            # if we load the pipeline code from the Hub
            # make sure to overwrite the `revison`
            revision = hub_revision

380
        return get_class_from_dynamic_module(
381
382
383
384
            custom_pipeline,
            module_file=file_name,
            class_name=class_name,
            cache_dir=cache_dir,
385
            revision=revision,
386
387
388
389
390
391
        )

    if class_obj != DiffusionPipeline:
        return class_obj

    diffusers_module = importlib.import_module(class_obj.__module__.split(".")[0])
392
    class_name = config["_class_name"]
393
    class_name = class_name[4:] if class_name.startswith("Flax") else class_name
394
395

    pipeline_cls = getattr(diffusers_module, class_name)
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410

    if load_connected_pipeline:
        from .auto_pipeline import _get_connected_pipeline

        connected_pipeline_cls = _get_connected_pipeline(pipeline_cls)
        if connected_pipeline_cls is not None:
            logger.info(
                f"Loading connected pipeline {connected_pipeline_cls.__name__} instead of {pipeline_cls.__name__} as specified via `load_connected_pipeline=True`"
            )
        else:
            logger.info(f"{pipeline_cls.__name__} has no connected pipeline class. Loading {pipeline_cls.__name__}.")

        pipeline_cls = connected_pipeline_cls or pipeline_cls

    return pipeline_cls
411
412


413
414
415
416
417
418
419
420
421
422
423
def load_sub_model(
    library_name: str,
    class_name: str,
    importable_classes: List[Any],
    pipelines: Any,
    is_pipeline_module: bool,
    pipeline_class: Any,
    torch_dtype: torch.dtype,
    provider: Any,
    sess_options: Any,
    device_map: Optional[Union[Dict[str, torch.device], str]],
424
425
426
    max_memory: Optional[Dict[Union[int, str], Union[int, str]]],
    offload_folder: Optional[Union[str, os.PathLike]],
    offload_state_dict: bool,
427
428
429
430
431
432
    model_variants: Dict[str, str],
    name: str,
    from_flax: bool,
    variant: str,
    low_cpu_mem_usage: bool,
    cached_folder: Union[str, os.PathLike],
433
    revision: str = None,
434
435
436
437
):
    """Helper method to load the module `name` from `library_name` and `class_name`"""
    # retrieve class candidates
    class_obj, class_candidates = get_class_obj_and_candidates(
438
439
440
441
442
443
444
        library_name,
        class_name,
        importable_classes,
        pipelines,
        is_pipeline_module,
        component_name=name,
        cache_dir=cached_folder,
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
    )

    load_method_name = None
    # retrive load method name
    for class_name, class_candidate in class_candidates.items():
        if class_candidate is not None and issubclass(class_obj, class_candidate):
            load_method_name = importable_classes[class_name][1]

    # if load method name is None, then we have a dummy module -> raise Error
    if load_method_name is None:
        none_module = class_obj.__module__
        is_dummy_path = none_module.startswith(DUMMY_MODULES_FOLDER) or none_module.startswith(
            TRANSFORMERS_DUMMY_MODULES_FOLDER
        )
        if is_dummy_path and "dummy" in none_module:
            # call class_obj for nice error message of missing requirements
            class_obj()

        raise ValueError(
            f"The component {class_obj} of {pipeline_class} cannot be loaded as it does not seem to have"
            f" any of the loading methods defined in {ALL_IMPORTABLE_CLASSES}."
        )

    load_method = getattr(class_obj, load_method_name)

    # add kwargs to loading method
471
    diffusers_module = importlib.import_module(__name__.split(".")[0])
472
473
474
    loading_kwargs = {}
    if issubclass(class_obj, torch.nn.Module):
        loading_kwargs["torch_dtype"] = torch_dtype
475
    if issubclass(class_obj, diffusers_module.OnnxRuntimeModel):
476
477
478
        loading_kwargs["provider"] = provider
        loading_kwargs["sess_options"] = sess_options

479
    is_diffusers_model = issubclass(class_obj, diffusers_module.ModelMixin)
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496

    if is_transformers_available():
        transformers_version = version.parse(version.parse(transformers.__version__).base_version)
    else:
        transformers_version = "N/A"

    is_transformers_model = (
        is_transformers_available()
        and issubclass(class_obj, PreTrainedModel)
        and transformers_version >= version.parse("4.20.0")
    )

    # When loading a transformers model, if the device_map is None, the weights will be initialized as opposed to diffusers.
    # To make default loading faster we set the `low_cpu_mem_usage=low_cpu_mem_usage` flag which is `True` by default.
    # This makes sure that the weights won't be initialized which significantly speeds up loading.
    if is_diffusers_model or is_transformers_model:
        loading_kwargs["device_map"] = device_map
497
498
499
        loading_kwargs["max_memory"] = max_memory
        loading_kwargs["offload_folder"] = offload_folder
        loading_kwargs["offload_state_dict"] = offload_state_dict
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
        loading_kwargs["variant"] = model_variants.pop(name, None)
        if from_flax:
            loading_kwargs["from_flax"] = True

        # the following can be deleted once the minimum required `transformers` version
        # is higher than 4.27
        if (
            is_transformers_model
            and loading_kwargs["variant"] is not None
            and transformers_version < version.parse("4.27.0")
        ):
            raise ImportError(
                f"When passing `variant='{variant}'`, please make sure to upgrade your `transformers` version to at least 4.27.0.dev0"
            )
        elif is_transformers_model and loading_kwargs["variant"] is None:
            loading_kwargs.pop("variant")

        # if `from_flax` and model is transformer model, can currently not load with `low_cpu_mem_usage`
        if not (from_flax and is_transformers_model):
            loading_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
        else:
            loading_kwargs["low_cpu_mem_usage"] = False

    # check if the module is in a subdirectory
    if os.path.isdir(os.path.join(cached_folder, name)):
        loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)
    else:
        # else load from the root directory
        loaded_sub_model = load_method(cached_folder, **loading_kwargs)

    return loaded_sub_model


533
class DiffusionPipeline(ConfigMixin, PushToHubMixin):
534
    r"""
Steven Liu's avatar
Steven Liu committed
535
    Base class for all pipelines.
536

Steven Liu's avatar
Steven Liu committed
537
538
    [`DiffusionPipeline`] stores all components (models, schedulers, and processors) for diffusion pipelines and
    provides methods for loading, downloading and saving models. It also includes methods to:
539
540

        - move all PyTorch modules to the device of your choice
541
        - enable/disable the progress bar for the denoising iteration
542
543
544

    Class attributes:

Steven Liu's avatar
Steven Liu committed
545
546
        - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
          diffusion pipeline's components.
547
        - **_optional_components** (`List[str]`) -- List of all optional components that don't have to be passed to the
Steven Liu's avatar
Steven Liu committed
548
          pipeline to function (should be overridden by subclasses).
549
    """
550

551
    config_name = "model_index.json"
552
    model_cpu_offload_seq = None
553
    _optional_components = []
554
    _exclude_from_cpu_offload = []
555
    _load_connected_pipes = False
556
    _is_onnx = False
557
558
559

    def register_modules(self, **kwargs):
        # import it here to avoid circular import
560
561
        diffusers_module = importlib.import_module(__name__.split(".")[0])
        pipelines = getattr(diffusers_module, "pipelines")
562
563
564

        for name, module in kwargs.items():
            # retrieve library
565
            if module is None or isinstance(module, (tuple, list)) and module[0] is None:
566
567
                register_dict = {name: (None, None)}
            else:
568
                # register the config from the original module, not the dynamo compiled one
569
                not_compiled_module = _unwrap_model(module)
570

571
                library = not_compiled_module.__module__.split(".")[0]
572
573

                # check if the module is a pipeline module
574
                module_path_items = not_compiled_module.__module__.split(".")
575
576
                pipeline_dir = module_path_items[-2] if len(module_path_items) > 2 else None

577
                path = not_compiled_module.__module__.split(".")
578
579
580
581
582
                is_pipeline_module = pipeline_dir in path and hasattr(pipelines, pipeline_dir)

                # if library is not in LOADABLE_CLASSES, then it is a custom module.
                # Or if it's a pipeline module, then the module is inside the pipeline
                # folder so we set the library to module name.
583
                if is_pipeline_module:
584
                    library = pipeline_dir
585
                elif library not in LOADABLE_CLASSES:
586
                    library = not_compiled_module.__module__
587
588

                # retrieve class_name
589
                class_name = not_compiled_module.__class__.__name__
590
591
592
593
594
595
596
597
598

                register_dict = {name: (library, class_name)}

            # save model index config
            self.register_to_config(**register_dict)

            # set models
            setattr(self, name, module)

599
    def __setattr__(self, name: str, value: Any):
600
        if name in self.__dict__ and hasattr(self.config, name):
601
602
            # We need to overwrite the config if name exists in config
            if isinstance(getattr(self.config, name), (tuple, list)):
603
                if value is not None and self.config[name][0] is not None:
604
605
606
607
608
609
610
611
612
613
                    class_library_tuple = (value.__module__.split(".")[0], value.__class__.__name__)
                else:
                    class_library_tuple = (None, None)

                self.register_to_config(**{name: class_library_tuple})
            else:
                self.register_to_config(**{name: value})

        super().__setattr__(name, value)

614
615
616
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
617
        safe_serialization: bool = True,
618
        variant: Optional[str] = None,
619
620
        push_to_hub: bool = False,
        **kwargs,
621
622
    ):
        """
Steven Liu's avatar
Steven Liu committed
623
624
625
        Save all saveable variables of the pipeline to a directory. A pipeline variable can be saved and loaded if its
        class implements both a save and loading method. The pipeline is easily reloaded using the
        [`~DiffusionPipeline.from_pretrained`] class method.
626
627
628

        Arguments:
            save_directory (`str` or `os.PathLike`):
Steven Liu's avatar
Steven Liu committed
629
                Directory to save a pipeline to. Will be created if it doesn't exist.
630
            safe_serialization (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
631
                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
632
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
633
                If specified, weights are saved in the format `pytorch_model.<variant>.bin`.
634
635
636
637
638
639
            push_to_hub (`bool`, *optional*, defaults to `False`):
                Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the
                repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
                namespace).
            kwargs (`Dict[str, Any]`, *optional*):
                Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
640
641
        """
        model_index_dict = dict(self.config)
642
643
        model_index_dict.pop("_class_name", None)
        model_index_dict.pop("_diffusers_version", None)
644
        model_index_dict.pop("_module", None)
645
        model_index_dict.pop("_name_or_path", None)
646

647
648
649
650
651
652
653
654
        if push_to_hub:
            commit_message = kwargs.pop("commit_message", None)
            private = kwargs.pop("private", False)
            create_pr = kwargs.pop("create_pr", False)
            token = kwargs.pop("token", None)
            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
            repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id

655
656
657
658
659
660
661
662
663
664
665
666
667
668
        expected_modules, optional_kwargs = self._get_signature_keys(self)

        def is_saveable_module(name, value):
            if name not in expected_modules:
                return False
            if name in self._optional_components and value[0] is None:
                return False
            return True

        model_index_dict = {k: v for k, v in model_index_dict.items() if is_saveable_module(k, v)}
        for pipeline_component_name in model_index_dict.keys():
            sub_model = getattr(self, pipeline_component_name)
            model_cls = sub_model.__class__

669
670
671
            # Dynamo wraps the original model in a private class.
            # I didn't find a public API to get the original class.
            if is_compiled_module(sub_model):
672
                sub_model = _unwrap_model(sub_model)
673
674
                model_cls = sub_model.__class__

675
676
677
            save_method_name = None
            # search for the model's base class in LOADABLE_CLASSES
            for library_name, library_classes in LOADABLE_CLASSES.items():
678
679
680
681
682
683
684
                if library_name in sys.modules:
                    library = importlib.import_module(library_name)
                else:
                    logger.info(
                        f"{library_name} is not installed. Cannot save {pipeline_component_name} as {library_classes} from {library_name}"
                    )

685
686
687
688
689
690
691
692
693
                for base_class, save_load_methods in library_classes.items():
                    class_candidate = getattr(library, base_class, None)
                    if class_candidate is not None and issubclass(model_cls, class_candidate):
                        # if we found a suitable base class in LOADABLE_CLASSES then grab its save method
                        save_method_name = save_load_methods[0]
                        break
                if save_method_name is not None:
                    break

694
695
696
697
698
699
            if save_method_name is None:
                logger.warn(f"self.{pipeline_component_name}={sub_model} of type {type(sub_model)} cannot be saved.")
                # make sure that unsaveable components are not tried to be loaded afterward
                self.register_to_config(**{pipeline_component_name: (None, None)})
                continue

700
701
702
703
704
            save_method = getattr(sub_model, save_method_name)

            # Call the save method with the argument safe_serialization only if it's supported
            save_method_signature = inspect.signature(save_method)
            save_method_accept_safe = "safe_serialization" in save_method_signature.parameters
705
706
707
            save_method_accept_variant = "variant" in save_method_signature.parameters

            save_kwargs = {}
708
            if save_method_accept_safe:
709
710
711
712
713
                save_kwargs["safe_serialization"] = safe_serialization
            if save_method_accept_variant:
                save_kwargs["variant"] = variant

            save_method(os.path.join(save_directory, pipeline_component_name), **save_kwargs)
714

715
716
717
        # finally save the config
        self.save_config(save_directory)

718
719
720
721
722
723
724
725
726
        if push_to_hub:
            self._upload_folder(
                save_directory,
                repo_id,
                token=token,
                commit_message=commit_message,
                create_pr=create_pr,
            )

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 to(self, *args, **kwargs):
        r"""
        Performs Pipeline dtype and/or device conversion. A torch.dtype and torch.device are inferred from the
        arguments of `self.to(*args, **kwargs).`

        <Tip>

            If the pipeline already has the correct torch.dtype and torch.device, then it is returned as is. Otherwise,
            the returned pipeline is a copy of self with the desired torch.dtype and torch.device.

        </Tip>


        Here are the ways to call `to`:

        - `to(dtype, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified
          [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype)
        - `to(device, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified
          [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device)
        - `to(device=None, dtype=None, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the
          specified [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) and
          [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype)

        Arguments:
            dtype (`torch.dtype`, *optional*):
                Returns a pipeline with the specified
                [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype)
            device (`torch.Device`, *optional*):
                Returns a pipeline with the specified
                [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device)
            silence_dtype_warnings (`str`, *optional*, defaults to `False`):
                Whether to omit warnings if the target `dtype` is not compatible with the target `device`.

        Returns:
            [`DiffusionPipeline`]: The pipeline converted to specified `dtype` and/or `dtype`.
        """

        torch_dtype = kwargs.pop("torch_dtype", None)
        if torch_dtype is not None:
766
            deprecate("torch_dtype", "0.27.0", "")
767
768
        torch_device = kwargs.pop("torch_device", None)
        if torch_device is not None:
769
            deprecate("torch_device", "0.27.0", "")
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
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818

        dtype_kwarg = kwargs.pop("dtype", None)
        device_kwarg = kwargs.pop("device", None)
        silence_dtype_warnings = kwargs.pop("silence_dtype_warnings", False)

        if torch_dtype is not None and dtype_kwarg is not None:
            raise ValueError(
                "You have passed both `torch_dtype` and `dtype` as a keyword argument. Please make sure to only pass `dtype`."
            )

        dtype = torch_dtype or dtype_kwarg

        if torch_device is not None and device_kwarg is not None:
            raise ValueError(
                "You have passed both `torch_device` and `device` as a keyword argument. Please make sure to only pass `device`."
            )

        device = torch_device or device_kwarg

        dtype_arg = None
        device_arg = None
        if len(args) == 1:
            if isinstance(args[0], torch.dtype):
                dtype_arg = args[0]
            else:
                device_arg = torch.device(args[0]) if args[0] is not None else None
        elif len(args) == 2:
            if isinstance(args[0], torch.dtype):
                raise ValueError(
                    "When passing two arguments, make sure the first corresponds to `device` and the second to `dtype`."
                )
            device_arg = torch.device(args[0]) if args[0] is not None else None
            dtype_arg = args[1]
        elif len(args) > 2:
            raise ValueError("Please make sure to pass at most two arguments (`device` and `dtype`) `.to(...)`")

        if dtype is not None and dtype_arg is not None:
            raise ValueError(
                "You have passed `dtype` both as an argument and as a keyword argument. Please only pass one of the two."
            )

        dtype = dtype or dtype_arg

        if device is not None and device_arg is not None:
            raise ValueError(
                "You have passed `device` both as an argument and as a keyword argument. Please only pass one of the two."
            )

        device = device or device_arg
819

820
821
822
823
824
        # throw warning if pipeline is in "offloaded"-mode but user tries to manually set to GPU.
        def module_is_sequentially_offloaded(module):
            if not is_accelerate_available() or is_accelerate_version("<", "0.14.0"):
                return False

Patrick von Platen's avatar
Patrick von Platen committed
825
826
827
            return hasattr(module, "_hf_hook") and not isinstance(
                module._hf_hook, (accelerate.hooks.CpuOffload, accelerate.hooks.AlignDevicesHook)
            )
828
829
830
831
832
833
834
835
836
837
838

        def module_is_offloaded(module):
            if not is_accelerate_available() or is_accelerate_version("<", "0.17.0.dev0"):
                return False

            return hasattr(module, "_hf_hook") and isinstance(module._hf_hook, accelerate.hooks.CpuOffload)

        # .to("cuda") would raise an error if the pipeline is sequentially offloaded, so we raise our own to make it clearer
        pipeline_is_sequentially_offloaded = any(
            module_is_sequentially_offloaded(module) for _, module in self.components.items()
        )
839
        if pipeline_is_sequentially_offloaded and device and torch.device(device).type == "cuda":
840
841
842
843
844
845
            raise ValueError(
                "It seems like you have activated sequential model offloading by calling `enable_sequential_cpu_offload`, but are now attempting to move the pipeline to GPU. This is not compatible with offloading. Please, move your pipeline `.to('cpu')` or consider removing the move altogether if you use sequential offloading."
            )

        # Display a warning in this case (the operation succeeds but the benefits are lost)
        pipeline_is_offloaded = any(module_is_offloaded(module) for _, module in self.components.items())
846
        if pipeline_is_offloaded and device and torch.device(device).type == "cuda":
847
848
849
850
            logger.warning(
                f"It seems like you have activated model offloading by calling `enable_model_cpu_offload`, but are now manually moving the pipeline to GPU. It is strongly recommended against doing so as memory gains from offloading are likely to be lost. Offloading automatically takes care of moving the individual components {', '.join(self.components.keys())} to GPU when needed. To make sure offloading works as expected, you should consider moving the pipeline back to CPU: `pipeline.to('cpu')` or removing the move altogether if you use offloading."
            )

851
        module_names, _ = self._get_signature_keys(self)
852
853
        modules = [getattr(self, n, None) for n in module_names]
        modules = [m for m in modules if isinstance(m, torch.nn.Module)]
854

855
        is_offloaded = pipeline_is_offloaded or pipeline_is_sequentially_offloaded
856
        for module in modules:
Patrick von Platen's avatar
Patrick von Platen committed
857
858
            is_loaded_in_8bit = hasattr(module, "is_loaded_in_8bit") and module.is_loaded_in_8bit

859
            if is_loaded_in_8bit and dtype is not None:
Patrick von Platen's avatar
Patrick von Platen committed
860
861
862
863
                logger.warning(
                    f"The module '{module.__class__.__name__}' has been loaded in 8bit and conversion to {torch_dtype} is not yet supported. Module is still in 8bit precision."
                )

864
            if is_loaded_in_8bit and device is not None:
Patrick von Platen's avatar
Patrick von Platen committed
865
866
867
868
                logger.warning(
                    f"The module '{module.__class__.__name__}' has been loaded in 8bit and moving it to {torch_dtype} via `.to()` is not yet supported. Module is still on {module.device}."
                )
            else:
869
                module.to(device, dtype)
Patrick von Platen's avatar
Patrick von Platen committed
870

871
872
            if (
                module.dtype == torch.float16
873
                and str(device) in ["cpu"]
874
875
876
877
                and not silence_dtype_warnings
                and not is_offloaded
            ):
                logger.warning(
878
                    "Pipelines loaded with `dtype=torch.float16` cannot run with `cpu` device. It"
879
880
881
882
883
                    " is not recommended to move them to `cpu` as running them will fail. Please make"
                    " sure to use an accelerator to run the pipeline in inference, due to the lack of"
                    " support for`float16` operations on this device in PyTorch. Please, remove the"
                    " `torch_dtype=torch.float16` argument, or use another device for inference."
                )
884
885
886
887
888
889
890
891
        return self

    @property
    def device(self) -> torch.device:
        r"""
        Returns:
            `torch.device`: The torch device on which the pipeline is located.
        """
892
        module_names, _ = self._get_signature_keys(self)
893
894
895
896
897
        modules = [getattr(self, n, None) for n in module_names]
        modules = [m for m in modules if isinstance(m, torch.nn.Module)]

        for module in modules:
            return module.device
898

899
900
        return torch.device("cpu")

901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
    @property
    def dtype(self) -> torch.dtype:
        r"""
        Returns:
            `torch.dtype`: The torch dtype on which the pipeline is located.
        """
        module_names, _ = self._get_signature_keys(self)
        modules = [getattr(self, n, None) for n in module_names]
        modules = [m for m in modules if isinstance(m, torch.nn.Module)]

        for module in modules:
            return module.dtype

        return torch.float32

916
    @classmethod
917
    @validate_hf_hub_args
918
919
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
        r"""
Steven Liu's avatar
Steven Liu committed
920
        Instantiate a PyTorch diffusion pipeline from pretrained pipeline weights.
921

Steven Liu's avatar
Steven Liu committed
922
        The pipeline is set in evaluation mode (`model.eval()`) by default.
923

Steven Liu's avatar
Steven Liu committed
924
        If you get the error message below, you need to finetune the weights for your downstream task:
925

Steven Liu's avatar
Steven Liu committed
926
927
928
929
930
        ```
        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
        - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
        You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
        ```
931
932
933
934
935

        Parameters:
            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
                Can be either:

Steven Liu's avatar
Steven Liu committed
936
937
938
939
940
                    - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
                      hosted on the Hub.
                    - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
                      saved using
                    [`~DiffusionPipeline.save_pretrained`].
941
            torch_dtype (`str` or `torch.dtype`, *optional*):
Steven Liu's avatar
Steven Liu committed
942
943
                Override the default `torch.dtype` and load the model with another dtype. If "auto" is passed, the
                dtype is automatically derived from the model's weights.
944
945
946
947
            custom_pipeline (`str`, *optional*):

                <Tip warning={true}>

Steven Liu's avatar
Steven Liu committed
948
                🧪 This is an experimental feature and may change in the future.
949
950
951
952
953

                </Tip>

                Can be either:

Steven Liu's avatar
Steven Liu committed
954
955
956
                    - A string, the *repo id* (for example `hf-internal-testing/diffusers-dummy-pipeline`) of a custom
                      pipeline hosted on the Hub. The repository must contain a file called pipeline.py that defines
                      the custom pipeline.
957
                    - A string, the *file name* of a community pipeline hosted on GitHub under
Steven Liu's avatar
Steven Liu committed
958
959
960
961
962
963
                      [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file
                      names must match the file name and not the pipeline script (`clip_guided_stable_diffusion`
                      instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the
                      current main branch of GitHub.
                    - A path to a directory (`./my_pipeline_directory/`) containing a custom pipeline. The directory
                      must contain a file called `pipeline.py` that defines the custom pipeline.
964
965
966
967
968
969
970

                For more information on how to load and create custom pipelines, please have a look at [Loading and
                Adding Custom
                Pipelines](https://huggingface.co/docs/diffusers/using-diffusers/custom_pipeline_overview)
            force_download (`bool`, *optional*, defaults to `False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
971
            cache_dir (`Union[str, os.PathLike]`, *optional*):
Steven Liu's avatar
Steven Liu committed
972
973
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
974
            resume_download (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
975
976
                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
                incompletely downloaded files are deleted.
977
            proxies (`Dict[str, str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
978
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
979
980
981
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            output_loading_info(`bool`, *optional*, defaults to `False`):
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
Steven Liu's avatar
Steven Liu committed
982
983
984
            local_files_only (`bool`, *optional*, defaults to `False`):
                Whether to only load local model weights and configuration files or not. If set to `True`, the model
                won't be downloaded from the Hub.
985
            token (`str` or *bool*, *optional*):
Steven Liu's avatar
Steven Liu committed
986
987
                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                `diffusers-cli login` (stored in `~/.huggingface`) is used.
988
            revision (`str`, *optional*, defaults to `"main"`):
Steven Liu's avatar
Steven Liu committed
989
990
991
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            custom_revision (`str`, *optional*, defaults to `"main"`):
992
                The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
Steven Liu's avatar
Steven Liu committed
993
994
                `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
                custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
995
            mirror (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
996
997
998
                Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                information.
999
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
Steven Liu's avatar
Steven Liu committed
1000
1001
                A map that specifies where each submodule should go. It doesn’t need to be defined for each
                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
1002
1003
                same device.

Steven Liu's avatar
Steven Liu committed
1004
                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
1005
1006
                more information about each option see [designing a device
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
1007
            max_memory (`Dict`, *optional*):
Steven Liu's avatar
Steven Liu committed
1008
1009
                A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                each GPU and the available CPU RAM if unset.
1010
            offload_folder (`str` or `os.PathLike`, *optional*):
Steven Liu's avatar
Steven Liu committed
1011
                The path to offload weights if device_map contains the value `"disk"`.
1012
            offload_state_dict (`bool`, *optional*):
Steven Liu's avatar
Steven Liu committed
1013
1014
1015
                If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
                the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
                when there is some disk offload.
1016
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
Steven Liu's avatar
Steven Liu committed
1017
1018
1019
1020
                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
                argument to `True` will raise an error.
1021
            use_safetensors (`bool`, *optional*, defaults to `None`):
Steven Liu's avatar
Steven Liu committed
1022
1023
1024
                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
                weights. If set to `False`, safetensors weights are not loaded.
1025
1026
1027
1028
1029
            use_onnx (`bool`, *optional*, defaults to `None`):
                If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights
                will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is
                `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending
                with `.onnx` and `.pb`.
1030
            kwargs (remaining dictionary of keyword arguments, *optional*):
Steven Liu's avatar
Steven Liu committed
1031
1032
1033
                Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
                class). The overwritten components are passed directly to the pipelines `__init__` method. See example
                below for more information.
1034
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
1035
1036
                Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.
1037
1038
1039

        <Tip>

Steven Liu's avatar
Steven Liu committed
1040
1041
        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
        `huggingface-cli login`.
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064

        </Tip>

        Examples:

        ```py
        >>> from diffusers import DiffusionPipeline

        >>> # Download pipeline from huggingface.co and cache.
        >>> pipeline = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256")

        >>> # Download pipeline that requires an authorization token
        >>> # For more information on access tokens, please refer to this section
        >>> # of the documentation](https://huggingface.co/docs/hub/security-tokens)
        >>> pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")

        >>> # Use a different scheduler
        >>> from diffusers import LMSDiscreteScheduler

        >>> scheduler = LMSDiscreteScheduler.from_config(pipeline.scheduler.config)
        >>> pipeline.scheduler = scheduler
        ```
        """
1065
        cache_dir = kwargs.pop("cache_dir", None)
1066
1067
1068
        resume_download = kwargs.pop("resume_download", False)
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
1069
1070
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
1071
        revision = kwargs.pop("revision", None)
1072
        from_flax = kwargs.pop("from_flax", False)
1073
1074
1075
1076
1077
1078
        torch_dtype = kwargs.pop("torch_dtype", None)
        custom_pipeline = kwargs.pop("custom_pipeline", None)
        custom_revision = kwargs.pop("custom_revision", None)
        provider = kwargs.pop("provider", None)
        sess_options = kwargs.pop("sess_options", None)
        device_map = kwargs.pop("device_map", None)
1079
1080
1081
        max_memory = kwargs.pop("max_memory", None)
        offload_folder = kwargs.pop("offload_folder", None)
        offload_state_dict = kwargs.pop("offload_state_dict", False)
1082
        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
1083
        variant = kwargs.pop("variant", None)
1084
        use_safetensors = kwargs.pop("use_safetensors", None)
1085
        use_onnx = kwargs.pop("use_onnx", None)
1086
        load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
1087
1088
1089
1090

        # 1. Download the checkpoints and configs
        # use snapshot download here to get it working from from_pretrained
        if not os.path.isdir(pretrained_model_name_or_path):
Patrick von Platen's avatar
Patrick von Platen committed
1091
1092
1093
1094
1095
            if pretrained_model_name_or_path.count("/") > 1:
                raise ValueError(
                    f'The provided pretrained_model_name_or_path "{pretrained_model_name_or_path}"'
                    " is neither a valid local path nor a valid repo id. Please check the parameter."
                )
1096
            cached_folder = cls.download(
1097
1098
1099
1100
1101
1102
                pretrained_model_name_or_path,
                cache_dir=cache_dir,
                resume_download=resume_download,
                force_download=force_download,
                proxies=proxies,
                local_files_only=local_files_only,
1103
                token=token,
1104
                revision=revision,
1105
                from_flax=from_flax,
1106
                use_safetensors=use_safetensors,
1107
                use_onnx=use_onnx,
1108
                custom_pipeline=custom_pipeline,
1109
                custom_revision=custom_revision,
1110
                variant=variant,
1111
                load_connected_pipeline=load_connected_pipeline,
1112
                **kwargs,
1113
1114
1115
1116
            )
        else:
            cached_folder = pretrained_model_name_or_path

1117
1118
        config_dict = cls.load_config(cached_folder)

Patrick von Platen's avatar
Patrick von Platen committed
1119
1120
1121
        # pop out "_ignore_files" as it is only needed for download
        config_dict.pop("_ignore_files", None)

1122
1123
1124
        # 2. Define which model components should load variants
        # We retrieve the information by matching whether variant
        # model checkpoints exist in the subfolders
1125
1126
1127
1128
1129
        model_variants = {}
        if variant is not None:
            for folder in os.listdir(cached_folder):
                folder_path = os.path.join(cached_folder, folder)
                is_folder = os.path.isdir(folder_path) and folder in config_dict
1130
1131
1132
                variant_exists = is_folder and any(
                    p.split(".")[1].startswith(variant) for p in os.listdir(folder_path)
                )
1133
1134
1135
                if variant_exists:
                    model_variants[folder] = variant

1136
        # 3. Load the pipeline class, if using custom module then load it from the hub
1137
        # if we load from explicit class, let's use it
1138
1139
1140
1141
1142
1143
1144
1145
1146
        custom_class_name = None
        if os.path.isfile(os.path.join(cached_folder, f"{custom_pipeline}.py")):
            custom_pipeline = os.path.join(cached_folder, f"{custom_pipeline}.py")
        elif isinstance(config_dict["_class_name"], (list, tuple)) and os.path.isfile(
            os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py")
        ):
            custom_pipeline = os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py")
            custom_class_name = config_dict["_class_name"][1]

1147
        pipeline_class = _get_pipeline_class(
1148
1149
1150
1151
            cls,
            config_dict,
            load_connected_pipeline=load_connected_pipeline,
            custom_pipeline=custom_pipeline,
1152
            class_name=custom_class_name,
1153
1154
            cache_dir=cache_dir,
            revision=custom_revision,
1155
        )
1156

1157
        # DEPRECATED: To be removed in 1.0.0
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
        if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse(
            version.parse(config_dict["_diffusers_version"]).base_version
        ) <= version.parse("0.5.1"):
            from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy

            pipeline_class = StableDiffusionInpaintPipelineLegacy

            deprecation_message = (
                "You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the"
                f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For"
                " better inpainting results, we strongly suggest using Stable Diffusion's official inpainting"
                " checkpoint: https://huggingface.co/runwayml/stable-diffusion-inpainting instead or adapting your"
                f" checkpoint {pretrained_model_name_or_path} to the format of"
                " https://huggingface.co/runwayml/stable-diffusion-inpainting. Note that we do not actively maintain"
                " the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0."
            )
            deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False)

1176
1177
1178
        # 4. Define expected modules given pipeline signature
        # and define non-None initialized modules (=`init_kwargs`)

1179
1180
1181
1182
1183
1184
1185
1186
1187
        # some modules can be passed directly to the init
        # in this case they are already instantiated in `kwargs`
        # extract them here
        expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class)
        passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
        passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}

        init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs)

1188
1189
1190
1191
1192
1193
        # define init kwargs and make sure that optional component modules are filtered out
        init_kwargs = {
            k: init_dict.pop(k)
            for k in optional_kwargs
            if k in init_dict and k not in pipeline_class._optional_components
        }
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
        init_kwargs = {**init_kwargs, **passed_pipe_kwargs}

        # remove `null` components
        def load_module(name, value):
            if value[0] is None:
                return False
            if name in passed_class_obj and passed_class_obj[name] is None:
                return False
            return True

        init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)}

1206
1207
1208
1209
1210
1211
1212
1213
        # Special case: safety_checker must be loaded separately when using `from_flax`
        if from_flax and "safety_checker" in init_dict and "safety_checker" not in passed_class_obj:
            raise NotImplementedError(
                "The safety checker cannot be automatically loaded when loading weights `from_flax`."
                " Please, pass `safety_checker=None` to `from_pretrained`, and load the safety checker"
                " separately if you need it."
            )

1214
        # 5. Throw nice warnings / errors for fast accelerate loading
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
        if len(unused_kwargs) > 0:
            logger.warning(
                f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored."
            )

        if low_cpu_mem_usage and not is_accelerate_available():
            low_cpu_mem_usage = False
            logger.warning(
                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
                " install accelerate\n```\n."
            )

        if device_map is not None and not is_torch_version(">=", "1.9.0"):
            raise NotImplementedError(
                "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `device_map=None`."
            )

        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
            raise NotImplementedError(
                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
                " `low_cpu_mem_usage=False`."
            )

        if low_cpu_mem_usage is False and device_map is not None:
            raise ValueError(
                f"You cannot set `low_cpu_mem_usage` to False while using device_map={device_map} for loading and"
                " dispatching. Please make sure to set `low_cpu_mem_usage=True`."
            )

        # import it here to avoid circular import
        from diffusers import pipelines

1250
        # 6. Load each module in the pipeline
1251
        for name, (library_name, class_name) in logging.tqdm(init_dict.items(), desc="Loading pipeline components..."):
1252
            # 6.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names
1253
            class_name = class_name[4:] if class_name.startswith("Flax") else class_name
1254

1255
            # 6.2 Define all importable classes
1256
            is_pipeline_module = hasattr(pipelines, library_name)
1257
            importable_classes = ALL_IMPORTABLE_CLASSES
1258
1259
            loaded_sub_model = None

1260
            # 6.3 Use passed sub model or load class_name from library_name
1261
            if name in passed_class_obj:
1262
1263
1264
1265
1266
                # if the model is in a pipeline module, then we load it from the pipeline
                # check that passed_class_obj has correct parent class
                maybe_raise_or_warn(
                    library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module
                )
1267
1268
1269

                loaded_sub_model = passed_class_obj[name]
            else:
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
                # load sub model
                loaded_sub_model = load_sub_model(
                    library_name=library_name,
                    class_name=class_name,
                    importable_classes=importable_classes,
                    pipelines=pipelines,
                    is_pipeline_module=is_pipeline_module,
                    pipeline_class=pipeline_class,
                    torch_dtype=torch_dtype,
                    provider=provider,
                    sess_options=sess_options,
                    device_map=device_map,
1282
1283
1284
                    max_memory=max_memory,
                    offload_folder=offload_folder,
                    offload_state_dict=offload_state_dict,
1285
1286
1287
1288
1289
1290
                    model_variants=model_variants,
                    name=name,
                    from_flax=from_flax,
                    variant=variant,
                    low_cpu_mem_usage=low_cpu_mem_usage,
                    cached_folder=cached_folder,
1291
                    revision=revision,
1292
                )
1293
1294
1295
                logger.info(
                    f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}."
                )
1296
1297
1298

            init_kwargs[name] = loaded_sub_model  # UNet(...), # DiffusionSchedule(...)

1299
1300
1301
1302
1303
1304
1305
1306
1307
        if pipeline_class._load_connected_pipes and os.path.isfile(os.path.join(cached_folder, "README.md")):
            modelcard = ModelCard.load(os.path.join(cached_folder, "README.md"))
            connected_pipes = {prefix: getattr(modelcard.data, prefix, [None])[0] for prefix in CONNECTED_PIPES_KEYS}
            load_kwargs = {
                "cache_dir": cache_dir,
                "resume_download": resume_download,
                "force_download": force_download,
                "proxies": proxies,
                "local_files_only": local_files_only,
1308
                "token": token,
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
                "revision": revision,
                "torch_dtype": torch_dtype,
                "custom_pipeline": custom_pipeline,
                "custom_revision": custom_revision,
                "provider": provider,
                "sess_options": sess_options,
                "device_map": device_map,
                "max_memory": max_memory,
                "offload_folder": offload_folder,
                "offload_state_dict": offload_state_dict,
                "low_cpu_mem_usage": low_cpu_mem_usage,
                "variant": variant,
                "use_safetensors": use_safetensors,
            }
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334

            def get_connected_passed_kwargs(prefix):
                connected_passed_class_obj = {
                    k.replace(f"{prefix}_", ""): w for k, w in passed_class_obj.items() if k.split("_")[0] == prefix
                }
                connected_passed_pipe_kwargs = {
                    k.replace(f"{prefix}_", ""): w for k, w in passed_pipe_kwargs.items() if k.split("_")[0] == prefix
                }

                connected_passed_kwargs = {**connected_passed_class_obj, **connected_passed_pipe_kwargs}
                return connected_passed_kwargs

1335
            connected_pipes = {
1336
1337
1338
                prefix: DiffusionPipeline.from_pretrained(
                    repo_id, **load_kwargs.copy(), **get_connected_passed_kwargs(prefix)
                )
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
                for prefix, repo_id in connected_pipes.items()
                if repo_id is not None
            }

            for prefix, connected_pipe in connected_pipes.items():
                # add connected pipes to `init_kwargs` with <prefix>_<component_name>, e.g. "prior_text_encoder"
                init_kwargs.update(
                    {"_".join([prefix, name]): component for name, component in connected_pipe.components.items()}
                )

1349
        # 7. Potentially add passed objects if expected
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
        missing_modules = set(expected_modules) - set(init_kwargs.keys())
        passed_modules = list(passed_class_obj.keys())
        optional_modules = pipeline_class._optional_components
        if len(missing_modules) > 0 and missing_modules <= set(passed_modules + optional_modules):
            for module in missing_modules:
                init_kwargs[module] = passed_class_obj.get(module, None)
        elif len(missing_modules) > 0:
            passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - optional_kwargs
            raise ValueError(
                f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed."
            )

1362
        # 8. Instantiate the pipeline
1363
        model = pipeline_class(**init_kwargs)
1364
1365
1366

        # 9. Save where the model was instantiated from
        model.register_to_config(_name_or_path=pretrained_model_name_or_path)
1367
1368
        return model

1369
1370
1371
1372
    @property
    def name_or_path(self) -> str:
        return getattr(self.config, "_name_or_path", None)

1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
    @property
    def _execution_device(self):
        r"""
        Returns the device on which the pipeline's models will be executed. After calling
        [`~DiffusionPipeline.enable_sequential_cpu_offload`] the execution device can only be inferred from
        Accelerate's module hooks.
        """
        for name, model in self.components.items():
            if not isinstance(model, torch.nn.Module) or name in self._exclude_from_cpu_offload:
                continue

            if not hasattr(model, "_hf_hook"):
                return self.device
            for module in model.modules():
                if (
                    hasattr(module, "_hf_hook")
                    and hasattr(module._hf_hook, "execution_device")
                    and module._hf_hook.execution_device is not None
                ):
                    return torch.device(module._hf_hook.execution_device)
        return self.device

1395
    def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
1396
1397
1398
1399
1400
        r"""
        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
1401
1402
1403
1404
1405
1406
1407

        Arguments:
            gpu_id (`int`, *optional*):
                The ID of the accelerator that shall be used in inference. If not specified, it will default to 0.
            device (`torch.Device` or `str`, *optional*, defaults to "cuda"):
                The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will
                default to "cuda".
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
        """
        if self.model_cpu_offload_seq is None:
            raise ValueError(
                "Model CPU offload cannot be enabled because no `model_cpu_offload_seq` class attribute is set."
            )

        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
            from accelerate import cpu_offload_with_hook
        else:
            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")

1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
        torch_device = torch.device(device)
        device_index = torch_device.index

        if gpu_id is not None and device_index is not None:
            raise ValueError(
                f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}"
                f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}"
            )

        # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0
1429
        self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0)
1430
1431
1432

        device_type = torch_device.type
        device = torch.device(f"{device_type}:{self._offload_gpu_id}")
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444

        if self.device.type != "cpu":
            self.to("cpu", silence_dtype_warnings=True)
            device_mod = getattr(torch, self.device.type, None)
            if hasattr(device_mod, "empty_cache") and device_mod.is_available():
                device_mod.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)

        all_model_components = {k: v for k, v in self.components.items() if isinstance(v, torch.nn.Module)}

        self._all_hooks = []
        hook = None
        for model_str in self.model_cpu_offload_seq.split("->"):
1445
            model = all_model_components.pop(model_str, None)
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
            if not isinstance(model, torch.nn.Module):
                continue

            _, hook = cpu_offload_with_hook(model, device, prev_module_hook=hook)
            self._all_hooks.append(hook)

        # CPU offload models that are not in the seq chain unless they are explicitly excluded
        # these models will stay on CPU until maybe_free_model_hooks is called
        # some models cannot be in the seq chain because they are iteratively called, such as controlnet
        for name, model in all_model_components.items():
            if not isinstance(model, torch.nn.Module):
                continue

            if name in self._exclude_from_cpu_offload:
                model.to(device)
            else:
                _, hook = cpu_offload_with_hook(model, device)
                self._all_hooks.append(hook)

    def maybe_free_model_hooks(self):
        r"""
1467
1468
1469
1470
        Function that offloads all components, removes all model hooks that were added when using
        `enable_model_cpu_offload` and then applies them again. In case the model has not been offloaded this function
        is a no-op. Make sure to add this function to the end of the `__call__` function of your pipeline so that it
        functions correctly when applying enable_model_cpu_offload.
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
        """
        if not hasattr(self, "_all_hooks") or len(self._all_hooks) == 0:
            # `enable_model_cpu_offload` has not be called, so silently do nothing
            return

        for hook in self._all_hooks:
            # offload model and remove hook from model
            hook.offload()
            hook.remove()

        # make sure the model is in the same state as before calling it
        self.enable_model_cpu_offload()

1484
    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
1485
        r"""
1486
1487
1488
1489
        Offloads all models to CPU using 🤗 Accelerate, significantly reducing memory usage. When called, the state
        dicts of all `torch.nn.Module` components (except those in `self._exclude_from_cpu_offload`) are saved to CPU
        and then moved to `torch.device('meta')` and loaded to GPU only when their specific submodule has its `forward`
        method called. Offloading happens on a submodule basis. Memory savings are higher than with
1490
        `enable_model_cpu_offload`, but performance is lower.
1491
1492
1493
1494
1495
1496
1497

        Arguments:
            gpu_id (`int`, *optional*):
                The ID of the accelerator that shall be used in inference. If not specified, it will default to 0.
            device (`torch.Device` or `str`, *optional*, defaults to "cuda"):
                The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will
                default to "cuda".
1498
1499
1500
1501
1502
1503
        """
        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
            from accelerate import cpu_offload
        else:
            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")

1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
        torch_device = torch.device(device)
        device_index = torch_device.index

        if gpu_id is not None and device_index is not None:
            raise ValueError(
                f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}"
                f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}"
            )

        # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0
1514
        self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0)
1515
1516
1517

        device_type = torch_device.type
        device = torch.device(f"{device_type}:{self._offload_gpu_id}")
1518
1519
1520

        if self.device.type != "cpu":
            self.to("cpu", silence_dtype_warnings=True)
1521
1522
1523
            device_mod = getattr(torch, self.device.type, None)
            if hasattr(device_mod, "empty_cache") and device_mod.is_available():
                device_mod.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536

        for name, model in self.components.items():
            if not isinstance(model, torch.nn.Module):
                continue

            if name in self._exclude_from_cpu_offload:
                model.to(device)
            else:
                # make sure to offload buffers if not all high level weights
                # are of type nn.Module
                offload_buffers = len(model._parameters) > 0
                cpu_offload(model, device, offload_buffers=offload_buffers)

1537
    @classmethod
1538
    @validate_hf_hub_args
1539
1540
    def download(cls, pretrained_model_name, **kwargs) -> Union[str, os.PathLike]:
        r"""
Steven Liu's avatar
Steven Liu committed
1541
        Download and cache a PyTorch diffusion pipeline from pretrained pipeline weights.
1542
1543

        Parameters:
Steven Liu's avatar
Steven Liu committed
1544
            pretrained_model_name (`str` or `os.PathLike`, *optional*):
Steven Liu's avatar
Steven Liu committed
1545
                A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
Steven Liu's avatar
Steven Liu committed
1546
                hosted on the Hub.
1547
1548
1549
            custom_pipeline (`str`, *optional*):
                Can be either:

Steven Liu's avatar
Steven Liu committed
1550
                    - A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained
Steven Liu's avatar
Steven Liu committed
1551
1552
                      pipeline hosted on the Hub. The repository must contain a file called `pipeline.py` that defines
                      the custom pipeline.
1553
1554

                    - A string, the *file name* of a community pipeline hosted on GitHub under
Steven Liu's avatar
Steven Liu committed
1555
1556
1557
1558
                      [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file
                      names must match the file name and not the pipeline script (`clip_guided_stable_diffusion`
                      instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the
                      current `main` branch of GitHub.
1559

Steven Liu's avatar
Steven Liu committed
1560
1561
                    - A path to a *directory* (`./my_pipeline_directory/`) containing a custom pipeline. The directory
                      must contain a file called `pipeline.py` that defines the custom pipeline.
1562

Steven Liu's avatar
Steven Liu committed
1563
                <Tip warning={true}>
1564

Steven Liu's avatar
Steven Liu committed
1565
                🧪 This is an experimental feature and may change in the future.
1566

Steven Liu's avatar
Steven Liu committed
1567
                </Tip>
1568

Steven Liu's avatar
Steven Liu committed
1569
1570
                For more information on how to load and create custom pipelines, take a look at [How to contribute a
                community pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/contribute_pipeline).
1571
1572
1573
1574
1575

            force_download (`bool`, *optional*, defaults to `False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
            resume_download (`bool`, *optional*, defaults to `False`):
Steven Liu's avatar
Steven Liu committed
1576
                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
Steven Liu's avatar
Steven Liu committed
1577
                incompletely downloaded files are deleted.
1578
            proxies (`Dict[str, str]`, *optional*):
Steven Liu's avatar
Steven Liu committed
1579
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
1580
1581
1582
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            output_loading_info(`bool`, *optional*, defaults to `False`):
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
Steven Liu's avatar
Steven Liu committed
1583
1584
1585
            local_files_only (`bool`, *optional*, defaults to `False`):
                Whether to only load local model weights and configuration files or not. If set to `True`, the model
                won't be downloaded from the Hub.
1586
            token (`str` or *bool*, *optional*):
Steven Liu's avatar
Steven Liu committed
1587
1588
                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                `diffusers-cli login` (stored in `~/.huggingface`) is used.
1589
            revision (`str`, *optional*, defaults to `"main"`):
Steven Liu's avatar
Steven Liu committed
1590
1591
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
Steven Liu's avatar
Steven Liu committed
1592
            custom_revision (`str`, *optional*, defaults to `"main"`):
1593
                The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
Steven Liu's avatar
Steven Liu committed
1594
1595
                `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
                custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
1596
            mirror (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
1597
1598
1599
                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                information.
1600
            variant (`str`, *optional*):
Steven Liu's avatar
Steven Liu committed
1601
1602
                Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.
1603
1604
1605
1606
1607
1608
1609
1610
1611
            use_safetensors (`bool`, *optional*, defaults to `None`):
                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
                weights. If set to `False`, safetensors weights are not loaded.
            use_onnx (`bool`, *optional*, defaults to `False`):
                If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights
                will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is
                `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending
                with `.onnx` and `.pb`.
1612
1613
1614
1615
            trust_remote_code (`bool`, *optional*, defaults to `False`):
                Whether or not to allow for custom pipelines and components defined on the Hub in their own files. This
                option should only be set to `True` for repositories you trust and in which you have read the code, as
                it will execute code present on the Hub on your local machine.
Steven Liu's avatar
Steven Liu committed
1616
1617
1618
1619

        Returns:
            `os.PathLike`:
                A path to the downloaded pipeline.
1620
1621
1622

        <Tip>

Steven Liu's avatar
Steven Liu committed
1623
1624
        To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
        `huggingface-cli login`.
1625
1626
1627
1628

        </Tip>

        """
1629
        cache_dir = kwargs.pop("cache_dir", None)
1630
1631
1632
        resume_download = kwargs.pop("resume_download", False)
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
1633
1634
        local_files_only = kwargs.pop("local_files_only", None)
        token = kwargs.pop("token", None)
1635
1636
1637
        revision = kwargs.pop("revision", None)
        from_flax = kwargs.pop("from_flax", False)
        custom_pipeline = kwargs.pop("custom_pipeline", None)
1638
        custom_revision = kwargs.pop("custom_revision", None)
1639
        variant = kwargs.pop("variant", None)
1640
        use_safetensors = kwargs.pop("use_safetensors", None)
1641
        use_onnx = kwargs.pop("use_onnx", None)
1642
        load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
1643
        trust_remote_code = kwargs.pop("trust_remote_code", False)
1644
1645
1646

        allow_pickle = False
        if use_safetensors is None:
1647
            use_safetensors = True
1648
            allow_pickle = True
1649
1650
1651
1652

        allow_patterns = None
        ignore_patterns = None

1653
        model_info_call_error: Optional[Exception] = None
1654
1655
        if not local_files_only:
            try:
1656
                info = model_info(pretrained_model_name, token=token, revision=revision)
1657
            except (HTTPError, OfflineModeIsEnabled, requests.ConnectionError) as e:
1658
1659
                logger.warn(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.")
                local_files_only = True
1660
                model_info_call_error = e  # save error to reraise it if model is not cached locally
1661

1662
1663
1664
1665
1666
        if not local_files_only:
            config_file = hf_hub_download(
                pretrained_model_name,
                cls.config_name,
                cache_dir=cache_dir,
1667
                revision=revision,
1668
1669
1670
                proxies=proxies,
                force_download=force_download,
                resume_download=resume_download,
1671
                token=token,
1672
1673
1674
            )

            config_dict = cls._dict_from_json_file(config_file)
Patrick von Platen's avatar
Patrick von Platen committed
1675
1676
            ignore_filenames = config_dict.pop("_ignore_files", [])

1677
            # retrieve all folder_names that contain relevant files
1678
            folder_names = [k for k, v in config_dict.items() if isinstance(v, list) and k != "_class_name"]
1679

1680
            filenames = {sibling.rfilename for sibling in info.siblings}
1681
1682
            model_filenames, variant_filenames = variant_compatible_siblings(filenames, variant=variant)

1683
1684
1685
1686
1687
1688
1689
1690
            diffusers_module = importlib.import_module(__name__.split(".")[0])
            pipelines = getattr(diffusers_module, "pipelines")

            # optionally create a custom component <> custom file mapping
            custom_components = {}
            for component in folder_names:
                module_candidate = config_dict[component][0]

1691
                if module_candidate is None or not isinstance(module_candidate, str):
1692
1693
                    continue

1694
1695
                # We compute candidate file path on the Hub. Do not use `os.path.join`.
                candidate_file = f"{component}/{module_candidate}.py"
1696
1697
1698
1699
1700
1701
1702
1703

                if candidate_file in filenames:
                    custom_components[component] = module_candidate
                elif module_candidate not in LOADABLE_CLASSES and not hasattr(pipelines, module_candidate):
                    raise ValueError(
                        f"{candidate_file} as defined in `model_index.json` does not exist in {pretrained_model_name} and is not a module in 'diffusers/pipelines'."
                    )

1704
1705
1706
1707
            if len(variant_filenames) == 0 and variant is not None:
                deprecation_message = (
                    f"You are trying to load the model files of the `variant={variant}`, but no such modeling files are available."
                    f"The default model files: {model_filenames} will be loaded instead. Make sure to not load from `variant={variant}`"
1708
                    "if such variant modeling files are not available. Doing so will lead to an error in v0.24.0 as defaulting to non-variant"
1709
1710
                    "modeling files is deprecated."
                )
1711
                deprecate("no variant default", "0.24.0", deprecation_message, standard_warn=False)
1712

Patrick von Platen's avatar
Patrick von Platen committed
1713
1714
1715
1716
            # remove ignored filenames
            model_filenames = set(model_filenames) - set(ignore_filenames)
            variant_filenames = set(variant_filenames) - set(ignore_filenames)

1717
1718
1719
            # if the whole pipeline is cached we don't have to ping the Hub
            if revision in DEPRECATED_REVISION_ARGS and version.parse(
                version.parse(__version__).base_version
Patrick von Platen's avatar
Patrick von Platen committed
1720
            ) >= version.parse("0.22.0"):
1721
                warn_deprecated_model_variant(pretrained_model_name, token, variant, revision, model_filenames)
1722

1723
            model_folder_names = {os.path.split(f)[0] for f in model_filenames if os.path.split(f)[0] in folder_names}
1724

1725
1726
1727
1728
1729
            custom_class_name = None
            if custom_pipeline is None and isinstance(config_dict["_class_name"], (list, tuple)):
                custom_pipeline = config_dict["_class_name"][0]
                custom_class_name = config_dict["_class_name"][1]

1730
1731
1732
1733
1734
            # all filenames compatible with variant will be added
            allow_patterns = list(model_filenames)

            # allow all patterns from non-model folders
            # this enables downloading schedulers, tokenizers, ...
1735
            allow_patterns += [f"{k}/*" for k in folder_names if k not in model_folder_names]
1736
1737
1738
1739
            # add custom component files
            allow_patterns += [f"{k}/{f}.py" for k, f in custom_components.items()]
            # add custom pipeline file
            allow_patterns += [f"{custom_pipeline}.py"] if f"{custom_pipeline}.py" in filenames else []
1740
            # also allow downloading config.json files with the model
1741
            allow_patterns += [os.path.join(k, "config.json") for k in model_folder_names]
1742
1743
1744
1745
1746
1747
1748
1749

            allow_patterns += [
                SCHEDULER_CONFIG_NAME,
                CONFIG_NAME,
                cls.config_name,
                CUSTOM_PIPELINE_FILE_NAME,
            ]

1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
            load_pipe_from_hub = custom_pipeline is not None and f"{custom_pipeline}.py" in filenames
            load_components_from_hub = len(custom_components) > 0

            if load_pipe_from_hub and not trust_remote_code:
                raise ValueError(
                    f"The repository for {pretrained_model_name} contains custom code in {custom_pipeline}.py which must be executed to correctly "
                    f"load the model. You can inspect the repository content at https://hf.co/{pretrained_model_name}/blob/main/{custom_pipeline}.py.\n"
                    f"Please pass the argument `trust_remote_code=True` to allow custom code to be run."
                )

            if load_components_from_hub and not trust_remote_code:
                raise ValueError(
                    f"The repository for {pretrained_model_name} contains custom code in {'.py, '.join([os.path.join(k, v) for k,v in custom_components.items()])} which must be executed to correctly "
                    f"load the model. You can inspect the repository content at {', '.join([f'https://hf.co/{pretrained_model_name}/{k}/{v}.py' for k,v in custom_components.items()])}.\n"
                    f"Please pass the argument `trust_remote_code=True` to allow custom code to be run."
                )

1767
1768
            # retrieve passed components that should not be downloaded
            pipeline_class = _get_pipeline_class(
1769
1770
1771
1772
                cls,
                config_dict,
                load_connected_pipeline=load_connected_pipeline,
                custom_pipeline=custom_pipeline,
1773
1774
1775
                repo_id=pretrained_model_name if load_pipe_from_hub else None,
                hub_revision=revision,
                class_name=custom_class_name,
1776
1777
                cache_dir=cache_dir,
                revision=custom_revision,
1778
1779
1780
1781
            )
            expected_components, _ = cls._get_signature_keys(pipeline_class)
            passed_components = [k for k in expected_components if k in kwargs]

1782
1783
1784
            if (
                use_safetensors
                and not allow_pickle
1785
1786
1787
                and not is_safetensors_compatible(
                    model_filenames, variant=variant, passed_components=passed_components
                )
1788
1789
            ):
                raise EnvironmentError(
1790
                    f"Could not find the necessary `safetensors` weights in {model_filenames} (variant={variant})"
1791
                )
1792
1793
            if from_flax:
                ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb"]
1794
1795
1796
            elif use_safetensors and is_safetensors_compatible(
                model_filenames, variant=variant, passed_components=passed_components
            ):
1797
1798
                ignore_patterns = ["*.bin", "*.msgpack"]

1799
1800
1801
1802
                use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx
                if not use_onnx:
                    ignore_patterns += ["*.onnx", "*.pb"]

1803
1804
                safetensors_variant_filenames = {f for f in variant_filenames if f.endswith(".safetensors")}
                safetensors_model_filenames = {f for f in model_filenames if f.endswith(".safetensors")}
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
                if (
                    len(safetensors_variant_filenames) > 0
                    and safetensors_model_filenames != safetensors_variant_filenames
                ):
                    logger.warn(
                        f"\nA mixture of {variant} and non-{variant} filenames will be loaded.\nLoaded {variant} filenames:\n[{', '.join(safetensors_variant_filenames)}]\nLoaded non-{variant} filenames:\n[{', '.join(safetensors_model_filenames - safetensors_variant_filenames)}\nIf this behavior is not expected, please check your folder structure."
                    )
            else:
                ignore_patterns = ["*.safetensors", "*.msgpack"]

1815
1816
1817
1818
                use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx
                if not use_onnx:
                    ignore_patterns += ["*.onnx", "*.pb"]

1819
1820
                bin_variant_filenames = {f for f in variant_filenames if f.endswith(".bin")}
                bin_model_filenames = {f for f in model_filenames if f.endswith(".bin")}
1821
1822
1823
1824
1825
                if len(bin_variant_filenames) > 0 and bin_model_filenames != bin_variant_filenames:
                    logger.warn(
                        f"\nA mixture of {variant} and non-{variant} filenames will be loaded.\nLoaded {variant} filenames:\n[{', '.join(bin_variant_filenames)}]\nLoaded non-{variant} filenames:\n[{', '.join(bin_model_filenames - bin_variant_filenames)}\nIf this behavior is not expected, please check your folder structure."
                    )

1826
1827
1828
1829
            # Don't download any objects that are passed
            allow_patterns = [
                p for p in allow_patterns if not (len(p.split("/")) == 2 and p.split("/")[0] in passed_components)
            ]
1830
1831
1832
1833

            if pipeline_class._load_connected_pipes:
                allow_patterns.append("README.md")

1834
1835
1836
            # Don't download index files of forbidden patterns either
            ignore_patterns = ignore_patterns + [f"{i}.index.*json" for i in ignore_patterns]

1837
1838
1839
1840
1841
            re_ignore_pattern = [re.compile(fnmatch.translate(p)) for p in ignore_patterns]
            re_allow_pattern = [re.compile(fnmatch.translate(p)) for p in allow_patterns]

            expected_files = [f for f in filenames if not any(p.match(f) for p in re_ignore_pattern)]
            expected_files = [f for f in expected_files if any(p.match(f) for p in re_allow_pattern)]
1842

1843
1844
            snapshot_folder = Path(config_file).parent
            pipeline_is_cached = all((snapshot_folder / f).is_file() for f in expected_files)
1845

1846
            if pipeline_is_cached and not force_download:
1847
1848
1849
                # if the pipeline is cached, we can directly return it
                # else call snapshot_download
                return snapshot_folder
1850

1851
1852
1853
        user_agent = {"pipeline_class": cls.__name__}
        if custom_pipeline is not None and not custom_pipeline.endswith(".py"):
            user_agent["custom_pipeline"] = custom_pipeline
1854
1855

        # download all allow_patterns - ignore_patterns
1856
        try:
1857
            cached_folder = snapshot_download(
1858
1859
1860
1861
1862
                pretrained_model_name,
                cache_dir=cache_dir,
                resume_download=resume_download,
                proxies=proxies,
                local_files_only=local_files_only,
1863
                token=token,
1864
1865
1866
1867
1868
                revision=revision,
                allow_patterns=allow_patterns,
                ignore_patterns=ignore_patterns,
                user_agent=user_agent,
            )
1869

1870
1871
            # retrieve pipeline class from local file
            cls_name = cls.load_config(os.path.join(cached_folder, "model_index.json")).get("_class_name", None)
1872
            cls_name = cls_name[4:] if isinstance(cls_name, str) and cls_name.startswith("Flax") else cls_name
1873

1874
1875
            diffusers_module = importlib.import_module(__name__.split(".")[0])
            pipeline_class = getattr(diffusers_module, cls_name, None) if isinstance(cls_name, str) else None
1876
1877

            if pipeline_class is not None and pipeline_class._load_connected_pipes:
1878
1879
1880
                modelcard = ModelCard.load(os.path.join(cached_folder, "README.md"))
                connected_pipes = sum([getattr(modelcard.data, k, []) for k in CONNECTED_PIPES_KEYS], [])
                for connected_pipe_repo_id in connected_pipes:
1881
1882
1883
1884
1885
1886
                    download_kwargs = {
                        "cache_dir": cache_dir,
                        "resume_download": resume_download,
                        "force_download": force_download,
                        "proxies": proxies,
                        "local_files_only": local_files_only,
1887
                        "token": token,
1888
1889
1890
1891
                        "variant": variant,
                        "use_safetensors": use_safetensors,
                    }
                    DiffusionPipeline.download(connected_pipe_repo_id, **download_kwargs)
1892
1893
1894

            return cached_folder

1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
        except FileNotFoundError:
            # Means we tried to load pipeline with `local_files_only=True` but the files have not been found in local cache.
            # This can happen in two cases:
            # 1. If the user passed `local_files_only=True`                    => we raise the error directly
            # 2. If we forced `local_files_only=True` when `model_info` failed => we raise the initial error
            if model_info_call_error is None:
                # 1. user passed `local_files_only=True`
                raise
            else:
                # 2. we forced `local_files_only=True` when `model_info` failed
                raise EnvironmentError(
                    f"Cannot load model {pretrained_model_name}: model is not cached locally and an error occured"
                    " while trying to fetch metadata from the Hub. Please check out the root cause in the stacktrace"
                    " above."
                ) from model_info_call_error
1910

1911
1912
    @classmethod
    def _get_signature_keys(cls, obj):
1913
1914
1915
        parameters = inspect.signature(obj.__init__).parameters
        required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty}
        optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty})
1916
        expected_modules = set(required_parameters.keys()) - {"self"}
1917
1918
1919
1920
1921
1922
1923

        optional_names = list(optional_parameters)
        for name in optional_names:
            if name in cls._optional_components:
                expected_modules.add(name)
                optional_parameters.remove(name)

1924
1925
1926
1927
1928
1929
        return expected_modules, optional_parameters

    @property
    def components(self) -> Dict[str, Any]:
        r"""
        The `self.components` property can be useful to run different pipelines with the same weights and
Steven Liu's avatar
Steven Liu committed
1930
1931
1932
1933
        configurations without reallocating additional memory.

        Returns (`dict`):
            A dictionary containing all the modules needed to initialize the pipeline.
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956

        Examples:

        ```py
        >>> from diffusers import (
        ...     StableDiffusionPipeline,
        ...     StableDiffusionImg2ImgPipeline,
        ...     StableDiffusionInpaintPipeline,
        ... )

        >>> text2img = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
        >>> img2img = StableDiffusionImg2ImgPipeline(**text2img.components)
        >>> inpaint = StableDiffusionInpaintPipeline(**text2img.components)
        ```
        """
        expected_modules, optional_parameters = self._get_signature_keys(self)
        components = {
            k: getattr(self, k) for k in self.config.keys() if not k.startswith("_") and k not in optional_parameters
        }

        if set(components.keys()) != expected_modules:
            raise ValueError(
                f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected"
1957
                f" {expected_modules} to be defined, but {components.keys()} are defined."
1958
1959
1960
1961
1962
1963
1964
            )

        return components

    @staticmethod
    def numpy_to_pil(images):
        """
Steven Liu's avatar
Steven Liu committed
1965
        Convert a NumPy image or a batch of images to a PIL image.
1966
        """
Patrick von Platen's avatar
Patrick von Platen committed
1967
        return numpy_to_pil(images)
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986

    def progress_bar(self, iterable=None, total=None):
        if not hasattr(self, "_progress_bar_config"):
            self._progress_bar_config = {}
        elif not isinstance(self._progress_bar_config, dict):
            raise ValueError(
                f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}."
            )

        if iterable is not None:
            return tqdm(iterable, **self._progress_bar_config)
        elif total is not None:
            return tqdm(total=total, **self._progress_bar_config)
        else:
            raise ValueError("Either `total` or `iterable` has to be defined.")

    def set_progress_bar_config(self, **kwargs):
        self._progress_bar_config = kwargs

1987
    def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
1988
        r"""
1989
1990
1991
        Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/). When this
        option is enabled, you should observe lower GPU memory usage and a potential speed up during inference. Speed
        up during training is not guaranteed.
1992

Steven Liu's avatar
Steven Liu committed
1993
        <Tip warning={true}>
1994

Steven Liu's avatar
Steven Liu committed
1995
1996
1997
1998
        ⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes
        precedent.

        </Tip>
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018

        Parameters:
            attention_op (`Callable`, *optional*):
                Override the default `None` operator for use as `op` argument to the
                [`memory_efficient_attention()`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention)
                function of xFormers.

        Examples:

        ```py
        >>> import torch
        >>> from diffusers import DiffusionPipeline
        >>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp

        >>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16)
        >>> pipe = pipe.to("cuda")
        >>> pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
        >>> # Workaround for not accepting attention shape using VAE for Flash Attention
        >>> pipe.vae.enable_xformers_memory_efficient_attention(attention_op=None)
        ```
2019
        """
2020
        self.set_use_memory_efficient_attention_xformers(True, attention_op)
2021
2022
2023

    def disable_xformers_memory_efficient_attention(self):
        r"""
Steven Liu's avatar
Steven Liu committed
2024
        Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
2025
2026
2027
        """
        self.set_use_memory_efficient_attention_xformers(False)

2028
2029
2030
    def set_use_memory_efficient_attention_xformers(
        self, valid: bool, attention_op: Optional[Callable] = None
    ) -> None:
2031
2032
2033
2034
2035
        # Recursively walk through all the children.
        # Any children which exposes the set_use_memory_efficient_attention_xformers method
        # gets the message
        def fn_recursive_set_mem_eff(module: torch.nn.Module):
            if hasattr(module, "set_use_memory_efficient_attention_xformers"):
2036
                module.set_use_memory_efficient_attention_xformers(valid, attention_op)
2037
2038
2039
2040

            for child in module.children():
                fn_recursive_set_mem_eff(child)

2041
2042
2043
        module_names, _ = self._get_signature_keys(self)
        modules = [getattr(self, n, None) for n in module_names]
        modules = [m for m in modules if isinstance(m, torch.nn.Module)]
2044

2045
2046
        for module in modules:
            fn_recursive_set_mem_eff(module)
2047
2048
2049

    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
        r"""
2050
        Enable sliced attention computation. When this option is enabled, the attention module splits the input tensor
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
        in slices to compute attention in several steps. For more than one attention head, the computation is performed
        sequentially over each head. This is useful to save some memory in exchange for a small speed decrease.

        <Tip warning={true}>

        ⚠️ Don't enable attention slicing if you're already using `scaled_dot_product_attention` (SDPA) from PyTorch
        2.0 or xFormers. These attention computations are already very memory efficient so you won't need to enable
        this function. If you enable attention slicing with SDPA or xFormers, it can lead to serious slow downs!

        </Tip>
2061
2062
2063
2064

        Args:
            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
Alexander Pivovarov's avatar
Alexander Pivovarov committed
2065
                `"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is
2066
2067
                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
                must be a multiple of `slice_size`.
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084

        Examples:

        ```py
        >>> import torch
        >>> from diffusers import StableDiffusionPipeline

        >>> pipe = StableDiffusionPipeline.from_pretrained(
        ...     "runwayml/stable-diffusion-v1-5",
        ...     torch_dtype=torch.float16,
        ...     use_safetensors=True,
        ... )

        >>> prompt = "a photo of an astronaut riding a horse on mars"
        >>> pipe.enable_attention_slicing()
        >>> image = pipe(prompt).images[0]
        ```
2085
2086
2087
2088
2089
        """
        self.set_attention_slice(slice_size)

    def disable_attention_slicing(self):
        r"""
Steven Liu's avatar
Steven Liu committed
2090
2091
        Disable sliced attention computation. If `enable_attention_slicing` was previously called, attention is
        computed in one step.
2092
2093
2094
2095
2096
        """
        # set slice_size = `None` to disable `attention slicing`
        self.enable_attention_slicing(None)

    def set_attention_slice(self, slice_size: Optional[int]):
2097
2098
2099
        module_names, _ = self._get_signature_keys(self)
        modules = [getattr(self, n, None) for n in module_names]
        modules = [m for m in modules if isinstance(m, torch.nn.Module) and hasattr(m, "set_attention_slice")]
2100

2101
2102
        for module in modules:
            module.set_attention_slice(slice_size)