-
Notifications
You must be signed in to change notification settings - Fork 7k
Expand file tree
/
Copy pathpipeline_utils.py
More file actions
2389 lines (2041 loc) · 117 KB
/
pipeline_utils.py
File metadata and controls
2389 lines (2041 loc) · 117 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# coding=utf-8
# Copyright 2025 The HuggingFace Inc. team.
# 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.
import fnmatch
import importlib
import inspect
import os
import re
import sys
import types
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Callable, Dict, List, Union, get_args, get_origin, get_type_hints
import httpx
import numpy as np
import PIL.Image
import requests
import torch
from huggingface_hub import (
DDUFEntry,
ModelCard,
create_repo,
hf_hub_download,
model_info,
read_dduf_file,
snapshot_download,
)
from huggingface_hub.utils import HfHubHTTPError, OfflineModeIsEnabled, validate_hf_hub_args
from packaging import version
from tqdm.auto import tqdm
from typing_extensions import Self
from .. import __version__
from ..configuration_utils import ConfigMixin
from ..models import AutoencoderKL
from ..models.attention_processor import FusedAttnProcessor2_0
from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, ModelMixin
from ..quantizers import PipelineQuantizationConfig
from ..quantizers.bitsandbytes.utils import _check_bnb_status
from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
from ..utils import (
CONFIG_NAME,
DEPRECATED_REVISION_ARGS,
BaseOutput,
PushToHubMixin,
_get_detailed_type,
_is_valid_type,
deprecate,
is_accelerate_available,
is_accelerate_version,
is_bitsandbytes_version,
is_hpu_available,
is_torch_npu_available,
is_torch_version,
is_transformers_version,
logging,
numpy_to_pil,
)
from ..utils.distributed_utils import is_torch_dist_rank_zero
from ..utils.hub_utils import _check_legacy_sharding_variant_format, load_or_create_model_card, populate_model_card
from ..utils.torch_utils import empty_device_cache, get_device, is_compiled_module
if is_torch_npu_available():
import torch_npu # noqa: F401
from .pipeline_loading_utils import (
ALL_IMPORTABLE_CLASSES,
CONNECTED_PIPES_KEYS,
CUSTOM_PIPELINE_FILE_NAME,
LOADABLE_CLASSES,
_download_dduf_file,
_fetch_class_library_tuple,
_get_custom_components_and_folders,
_get_custom_pipeline_class,
_get_final_device_map,
_get_ignore_patterns,
_get_pipeline_class,
_identify_model_variants,
_maybe_raise_error_for_incorrect_transformers,
_maybe_raise_warning_for_inpainting,
_maybe_warn_for_wrong_component_in_quant_config,
_resolve_custom_pipeline_and_cls,
_unwrap_model,
_update_init_kwargs_with_connected_pipeline,
filter_model_files,
load_sub_model,
maybe_raise_or_warn,
variant_compatible_siblings,
warn_deprecated_model_variant,
)
if is_accelerate_available():
import accelerate
LIBRARIES = []
for library in LOADABLE_CLASSES:
LIBRARIES.append(library)
SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device(), "cpu"]
logger = logging.get_logger(__name__)
@dataclass
class ImagePipelineOutput(BaseOutput):
"""
Output class for image pipelines.
Args:
images (`List[PIL.Image.Image]` or `np.ndarray`)
List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
num_channels)`.
"""
images: list[PIL.Image.Image] | np.ndarray
@dataclass
class AudioPipelineOutput(BaseOutput):
"""
Output class for audio pipelines.
Args:
audios (`np.ndarray`)
List of denoised audio samples of a NumPy array of shape `(batch_size, num_channels, sample_rate)`.
"""
audios: np.ndarray
class DeprecatedPipelineMixin:
"""
A mixin that can be used to mark a pipeline as deprecated.
Pipelines inheriting from this mixin will raise a warning when instantiated, indicating that they are deprecated
and won't receive updates past the specified version. Tests will be skipped for pipelines that inherit from this
mixin.
Example usage:
```python
class MyDeprecatedPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
_last_supported_version = "0.20.0"
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
```
"""
# Override this in the inheriting class to specify the last version that will support this pipeline
_last_supported_version = None
def __init__(self, *args, **kwargs):
# Get the class name for the warning message
class_name = self.__class__.__name__
# Get the last supported version or use the current version if not specified
version_info = getattr(self.__class__, "_last_supported_version", __version__)
# Raise a warning that this pipeline is deprecated
logger.warning(
f"The {class_name} has been deprecated and will not receive bug fixes or feature updates after Diffusers version {version_info}. "
)
# Call the parent class's __init__ method
super().__init__(*args, **kwargs)
class DiffusionPipeline(ConfigMixin, PushToHubMixin):
r"""
Base class for all pipelines.
[`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:
- move all PyTorch modules to the device of your choice
- enable/disable the progress bar for the denoising iteration
Class attributes:
- **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
diffusion pipeline's components.
- **_optional_components** (`List[str]`) -- List of all optional components that don't have to be passed to the
pipeline to function (should be overridden by subclasses).
"""
config_name = "model_index.json"
model_cpu_offload_seq = None
hf_device_map = None
_optional_components = []
_exclude_from_cpu_offload = []
_load_connected_pipes = False
_is_onnx = False
def register_modules(self, **kwargs):
for name, module in kwargs.items():
# retrieve library
if module is None or isinstance(module, (tuple, list)) and module[0] is None:
register_dict = {name: (None, None)}
else:
library, class_name = _fetch_class_library_tuple(module)
register_dict = {name: (library, class_name)}
# save model index config
self.register_to_config(**register_dict)
# set models
setattr(self, name, module)
def __setattr__(self, name: str, value: Any):
if name in self.__dict__ and hasattr(self.config, name):
# We need to overwrite the config if name exists in config
if isinstance(getattr(self.config, name), (tuple, list)):
if value is not None and self.config[name][0] is not None:
class_library_tuple = _fetch_class_library_tuple(value)
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)
def save_pretrained(
self,
save_directory: str | os.PathLike,
safe_serialization: bool = True,
variant: str | None = None,
max_shard_size: int | str | None = None,
push_to_hub: bool = False,
use_flashpack: bool = False,
**kwargs,
):
"""
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.
Arguments:
save_directory (`str` or `os.PathLike`):
Directory to save a pipeline to. Will be created if it doesn't exist.
safe_serialization (`bool`, *optional*, defaults to `True`):
Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
variant (`str`, *optional*):
If specified, weights are saved in the format `pytorch_model.<variant>.bin`.
max_shard_size (`int` or `str`, defaults to `None`):
The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size
lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5GB"`).
If expressed as an integer, the unit is bytes. Note that this limit will be decreased after a certain
period of time (starting from Oct 2024) to allow users to upgrade to the latest version of `diffusers`.
This is to establish a common default size for this argument across different libraries in the Hugging
Face ecosystem (`transformers`, and `accelerate`, for example).
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.
"""
model_index_dict = dict(self.config)
model_index_dict.pop("_class_name", None)
model_index_dict.pop("_diffusers_version", None)
model_index_dict.pop("_module", None)
model_index_dict.pop("_name_or_path", None)
if push_to_hub:
commit_message = kwargs.pop("commit_message", None)
private = kwargs.pop("private", None)
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
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__
# 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):
sub_model = _unwrap_model(sub_model)
model_cls = sub_model.__class__
save_method_name = None
# search for the model's base class in LOADABLE_CLASSES
for library_name, library_classes in LOADABLE_CLASSES.items():
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}"
)
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
if save_method_name is None:
logger.warning(
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
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
save_method_accept_variant = "variant" in save_method_signature.parameters
save_method_accept_max_shard_size = "max_shard_size" in save_method_signature.parameters
save_method_accept_flashpack = "use_flashpack" in save_method_signature.parameters
save_method_accept_peft_format = "save_peft_format" in save_method_signature.parameters
save_kwargs = {}
if save_method_accept_safe:
save_kwargs["safe_serialization"] = safe_serialization
if save_method_accept_variant:
save_kwargs["variant"] = variant
if save_method_accept_max_shard_size and max_shard_size is not None:
# max_shard_size is expected to not be None in ModelMixin
save_kwargs["max_shard_size"] = max_shard_size
if save_method_accept_flashpack:
save_kwargs["use_flashpack"] = use_flashpack
if save_method_accept_peft_format:
# Set save_peft_format=False for transformers>=5.0.0 compatibility
# In transformers 5.0.0+, the default save_peft_format=True adds "base_model.model" prefix
# to adapter keys, but from_pretrained expects keys without this prefix
save_kwargs["save_peft_format"] = False
save_method(os.path.join(save_directory, pipeline_component_name), **save_kwargs)
# finally save the config
self.save_config(save_directory)
if push_to_hub:
# Create a new empty model card and eventually tag it
model_card = load_or_create_model_card(repo_id, token=token, is_pipeline=True)
model_card = populate_model_card(model_card)
model_card.save(os.path.join(save_directory, "README.md"))
self._upload_folder(
save_directory,
repo_id,
token=token,
commit_message=commit_message,
create_pr=create_pr,
)
def to(self, *args, **kwargs) -> Self:
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.
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`.
"""
dtype = kwargs.pop("dtype", None)
device = kwargs.pop("device", None)
silence_dtype_warnings = kwargs.pop("silence_dtype_warnings", False)
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
device_type = torch.device(device).type if device is not None else None
pipeline_has_bnb = any(any((_check_bnb_status(module))) for _, module in self.components.items())
# 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
_, _, is_loaded_in_8bit_bnb = _check_bnb_status(module)
# https://github.com/huggingface/accelerate/pull/3907
if is_loaded_in_8bit_bnb and (
is_bitsandbytes_version("<", "0.48.0") or is_accelerate_version("<", "1.13.0.dev0")
):
return False
return hasattr(module, "_hf_hook") and (
isinstance(module._hf_hook, accelerate.hooks.AlignDevicesHook)
or hasattr(module._hf_hook, "hooks")
and isinstance(module._hf_hook.hooks[0], accelerate.hooks.AlignDevicesHook)
)
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()
)
is_pipeline_device_mapped = self._is_pipeline_device_mapped()
if is_pipeline_device_mapped:
raise ValueError(
"It seems like you have activated a device mapping strategy on the pipeline which doesn't allow explicit device placement using `to()`. You can call `reset_device_map()` to remove the existing device map from the pipeline."
)
if device_type in ["cuda", "xpu"]:
if pipeline_is_sequentially_offloaded and not pipeline_has_bnb:
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."
)
# PR: https://github.com/huggingface/accelerate/pull/3223/
elif pipeline_has_bnb and is_accelerate_version("<", "1.1.0.dev0"):
raise ValueError(
"You are trying to call `.to('cuda')` on a pipeline that has models quantized with `bitsandbytes`. Your current `accelerate` installation does not support it. Please upgrade the installation."
)
# 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())
if pipeline_is_offloaded and device_type in ["cuda", "xpu"]:
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."
)
# Enable generic support for Intel Gaudi accelerator using GPU/HPU migration
if device_type == "hpu" and kwargs.pop("hpu_migration", True) and is_hpu_available():
os.environ["PT_HPU_GPU_MIGRATION"] = "1"
logger.debug("Environment variable set: PT_HPU_GPU_MIGRATION=1")
import habana_frameworks.torch # noqa: F401
# HPU hardware check
if not (hasattr(torch, "hpu") and torch.hpu.is_available()):
raise ValueError("You are trying to call `.to('hpu')` but HPU device is unavailable.")
os.environ["PT_HPU_MAX_COMPOUND_OP_SIZE"] = "1"
logger.debug("Environment variable set: PT_HPU_MAX_COMPOUND_OP_SIZE=1")
if dtype in (torch.bfloat16, None) and kwargs.pop("sdp_on_bf16", True):
if hasattr(torch._C, "_set_math_sdp_allow_fp16_bf16_reduction"):
torch._C._set_math_sdp_allow_fp16_bf16_reduction(True)
logger.warning(
"Enabled SDP with BF16 precision on HPU. To disable, please use `.to('hpu', sdp_on_bf16=False)`"
)
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)]
is_offloaded = pipeline_is_offloaded or pipeline_is_sequentially_offloaded
for module in modules:
_, is_loaded_in_4bit_bnb, is_loaded_in_8bit_bnb = _check_bnb_status(module)
is_group_offloaded = self._maybe_raise_error_if_group_offload_active(module=module)
if (is_loaded_in_4bit_bnb or is_loaded_in_8bit_bnb) and dtype is not None:
logger.warning(
f"The module '{module.__class__.__name__}' has been loaded in `bitsandbytes` {'4bit' if is_loaded_in_4bit_bnb else '8bit'} and conversion to {dtype} is not supported. Module is still in {'4bit' if is_loaded_in_4bit_bnb else '8bit'} precision."
)
if is_loaded_in_8bit_bnb and device is not None and is_bitsandbytes_version("<", "0.48.0"):
logger.warning(
f"The module '{module.__class__.__name__}' has been loaded in `bitsandbytes` 8bit and moving it to {device} via `.to()` is not supported. Module is still on {module.device}."
"You need to upgrade bitsandbytes to at least 0.48.0"
)
# Note: we also handle this at the ModelMixin level. The reason for doing it here too is that modeling
# components can be from outside diffusers too, but still have group offloading enabled.
if (
self._maybe_raise_error_if_group_offload_active(raise_error=False, module=module)
and device is not None
):
logger.warning(
f"The module '{module.__class__.__name__}' is group offloaded and moving it to {device} via `.to()` is not supported."
)
# This can happen for `transformer` models. CPU placement was added in
# https://github.com/huggingface/transformers/pull/33122. So, we guard this accordingly.
if is_loaded_in_4bit_bnb and device is not None and is_transformers_version(">", "4.44.0"):
module.to(device=device)
# added here https://github.com/huggingface/transformers/pull/43258
if (
is_loaded_in_8bit_bnb
and device is not None
and is_transformers_version(">", "4.58.0")
and is_bitsandbytes_version(">=", "0.48.0")
):
module.to(device=device)
elif not is_loaded_in_4bit_bnb and not is_loaded_in_8bit_bnb and not is_group_offloaded:
module.to(device, dtype)
if (
module.dtype == torch.float16
and str(device) in ["cpu"]
and not silence_dtype_warnings
and not is_offloaded
):
logger.warning(
"Pipelines loaded with `dtype=torch.float16` cannot run with `cpu` device. It"
" 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."
)
return self
@property
def device(self) -> torch.device:
r"""
Returns:
`torch.device`: The torch device 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.device
return torch.device("cpu")
@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
@classmethod
@validate_hf_hub_args
def from_pretrained(cls, pretrained_model_name_or_path: str | os.PathLike, **kwargs) -> Self:
r"""
Instantiate a PyTorch diffusion pipeline from pretrained pipeline weights.
The pipeline is set in evaluation mode (`model.eval()`) by default.
If you get the error message below, you need to finetune the weights for your downstream task:
```
Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/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.
```
Parameters:
pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
Can be either:
- 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`].
- A path to a *directory* (for example `./my_pipeline_directory/`) containing a dduf file
torch_dtype (`torch.dtype` or `dict[str, Union[str, torch.dtype]]`, *optional*):
Override the default `torch.dtype` and load the model with another dtype. To load submodels with
different dtype pass a `dict` (for example `{'transformer': torch.bfloat16, 'vae': torch.float16}`).
Set the default dtype for unspecified components with `default` (for example `{'transformer':
torch.bfloat16, 'default': torch.float16}`). If a component is not specified and no default is set,
`torch.float32` is used.
custom_pipeline (`str`, *optional*):
> [!WARNING] > 🧪 This is an experimental feature and may change in the future.
Can be either:
- 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.
- A string, the *file name* of a community pipeline hosted on GitHub under
[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.
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.
cache_dir (`Union[str, os.PathLike]`, *optional*):
Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
is not used.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
'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.
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.
token (`str` or *bool*, *optional*):
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.
revision (`str`, *optional*, defaults to `"main"`):
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*):
The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
`revision` when loading a custom pipeline from the Hub. Defaults to the latest stable 🤗 Diffusers
version.
mirror (`str`, *optional*):
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.
device_map (`str`, *optional*):
Strategy that dictates how the different components of a pipeline should be placed on available
devices. Currently, only "balanced" `device_map` is supported. Check out
[this](https://huggingface.co/docs/diffusers/main/en/tutorials/inference_with_big_models#device-placement)
to know more.
max_memory (`Dict`, *optional*):
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.
offload_folder (`str` or `os.PathLike`, *optional*):
The path to offload weights if device_map contains the value `"disk"`.
offload_state_dict (`bool`, *optional*):
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.
low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
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.
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 `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`.
kwargs (remaining dictionary of keyword arguments, *optional*):
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.
variant (`str`, *optional*):
Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
loading `from_flax`.
dduf_file(`str`, *optional*):
Load weights from the specified dduf file.
disable_mmap ('bool', *optional*, defaults to 'False'):
Whether to disable mmap when loading a Safetensors model. This option can perform better when the model
is on a network mount or hard drive, which may not handle the seeky-ness of mmap very well.
> [!TIP] > To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in
with `hf > auth login`.
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("stable-diffusion-v1-5/stable-diffusion-v1-5")
>>> # Use a different scheduler
>>> from diffusers import LMSDiscreteScheduler
>>> scheduler = LMSDiscreteScheduler.from_config(pipeline.scheduler.config)
>>> pipeline.scheduler = scheduler
```
"""
# Copy the kwargs to re-use during loading connected pipeline.
kwargs_copied = kwargs.copy()
cache_dir = kwargs.pop("cache_dir", None)
force_download = kwargs.pop("force_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", None)
token = kwargs.pop("token", None)
revision = kwargs.pop("revision", None)
from_flax = kwargs.pop("from_flax", False)
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)
provider_options = kwargs.pop("provider_options", None)
device_map = kwargs.pop("device_map", None)
max_memory = kwargs.pop("max_memory", None)
offload_folder = kwargs.pop("offload_folder", None)
offload_state_dict = kwargs.pop("offload_state_dict", None)
low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
variant = kwargs.pop("variant", None)
dduf_file = kwargs.pop("dduf_file", None)
use_safetensors = kwargs.pop("use_safetensors", None)
use_onnx = kwargs.pop("use_onnx", None)
load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
quantization_config = kwargs.pop("quantization_config", None)
use_flashpack = kwargs.pop("use_flashpack", False)
disable_mmap = kwargs.pop("disable_mmap", False)
trust_remote_code = kwargs.pop("trust_remote_code", False)
if torch_dtype is not None and not isinstance(torch_dtype, dict) and not isinstance(torch_dtype, torch.dtype):
torch_dtype = torch.float32
logger.warning(
f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`."
)
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 quantization_config is not None and not isinstance(quantization_config, PipelineQuantizationConfig):
raise ValueError("`quantization_config` must be an instance of `PipelineQuantizationConfig`.")
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 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 device_map is not None and not is_accelerate_available():
raise NotImplementedError(
"Using `device_map` requires the `accelerate` library. Please install it using: `pip install accelerate`."
)
if device_map is not None and not isinstance(device_map, str):
raise ValueError("`device_map` must be a string.")
if device_map is not None and device_map not in SUPPORTED_DEVICE_MAP:
raise NotImplementedError(
f"{device_map} not supported. Supported strategies are: {', '.join(SUPPORTED_DEVICE_MAP)}"
)
if device_map is not None and device_map in SUPPORTED_DEVICE_MAP:
if is_accelerate_version("<", "0.28.0"):
raise NotImplementedError("Device placement requires `accelerate` version `0.28.0` or later.")
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`."
)
if dduf_file:
if custom_pipeline:
raise NotImplementedError("Custom pipelines are not supported with DDUF at the moment.")
if load_connected_pipeline:
raise NotImplementedError("Connected pipelines are not supported with DDUF at the moment.")
# 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):
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."
)
cached_folder = cls.download(
pretrained_model_name_or_path,
cache_dir=cache_dir,
force_download=force_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
from_flax=from_flax,
use_safetensors=use_safetensors,
use_onnx=use_onnx,
custom_pipeline=custom_pipeline,
custom_revision=custom_revision,
variant=variant,
dduf_file=dduf_file,
load_connected_pipeline=load_connected_pipeline,
trust_remote_code=trust_remote_code,
**kwargs,
)
else:
cached_folder = pretrained_model_name_or_path
# The variant filenames can have the legacy sharding checkpoint format that we check and throw
# a warning if detected.
if variant is not None and _check_legacy_sharding_variant_format(folder=cached_folder, variant=variant):
warn_msg = (
f"Warning: The repository contains sharded checkpoints for variant '{variant}' maybe in a deprecated format. "
"Please check your files carefully:\n\n"
"- Correct format example: diffusion_pytorch_model.fp16-00003-of-00003.safetensors\n"
"- Deprecated format example: diffusion_pytorch_model-00001-of-00002.fp16.safetensors\n\n"
"If you find any files in the deprecated format:\n"
"1. Remove all existing checkpoint files for this variant.\n"
"2. Re-obtain the correct files by running `save_pretrained()`.\n\n"
"This will ensure you're using the most up-to-date and compatible checkpoint format."
)
logger.warning(warn_msg)
dduf_entries = None
if dduf_file:
dduf_file_path = os.path.join(cached_folder, dduf_file)
dduf_entries = read_dduf_file(dduf_file_path)
# The reader contains already all the files needed, no need to check it again
cached_folder = ""
config_dict = cls.load_config(cached_folder, dduf_entries=dduf_entries)
if dduf_file:
_maybe_raise_error_for_incorrect_transformers(config_dict)
# pop out "_ignore_files" as it is only needed for download
config_dict.pop("_ignore_files", None)
# 2. Define which model components should load variants
# We retrieve the information by matching whether variant model checkpoints exist in the subfolders.
# Example: `diffusion_pytorch_model.safetensors` -> `diffusion_pytorch_model.fp16.safetensors`
# with variant being `"fp16"`.
model_variants = _identify_model_variants(folder=cached_folder, variant=variant, config=config_dict)
if len(model_variants) == 0 and variant is not None:
error_message = f"You are trying to load the model files of the `variant={variant}`, but no such modeling files are available."
raise ValueError(error_message)
# 3. Load the pipeline class, if using custom module then load it from the hub
# if we load from explicit class, let's use it
custom_pipeline, custom_class_name = _resolve_custom_pipeline_and_cls(
folder=cached_folder, config=config_dict, custom_pipeline=custom_pipeline
)
pipeline_class = _get_pipeline_class(
cls,
config=config_dict,
load_connected_pipeline=load_connected_pipeline,
custom_pipeline=custom_pipeline,
class_name=custom_class_name,
cache_dir=cache_dir,
revision=custom_revision,
trust_remote_code=trust_remote_code,
)
if device_map is not None and pipeline_class._load_connected_pipes:
raise NotImplementedError("`device_map` is not yet supported for connected pipelines.")
# DEPRECATED: To be removed in 1.0.0
# we are deprecating the `StableDiffusionInpaintPipelineLegacy` pipeline which gets loaded
# when a user requests for a `StableDiffusionInpaintPipeline` with `diffusers` version being <= 0.5.1.
_maybe_raise_warning_for_inpainting(
pipeline_class=pipeline_class,
pretrained_model_name_or_path=pretrained_model_name_or_path,
config=config_dict,
)
# 4. Define expected modules given pipeline signature
# and define non-None initialized modules (=`init_kwargs`)
# 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)
expected_types = pipeline_class._get_signature_types()
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)
# 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
}
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)}
# 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."
)
# 5. Throw nice warnings / errors for fast accelerate loading
if len(unused_kwargs) > 0:
logger.warning(
f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored."
)
# import it here to avoid circular import
from diffusers import pipelines
# 6. device map delegation
final_device_map = None
if device_map is not None:
final_device_map = _get_final_device_map(
device_map=device_map,
pipeline_class=pipeline_class,