# coding=utf-8

# Copyright 2023 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 warnings

from dataclasses import dataclass

from pathlib import Path

from typing import Any, Callable, Dict, List, Optional, Union

import numpy as np

import PIL

import torch

from huggingface_hub import ModelCard, create_repo, hf_hub_download, model_info, snapshot_download

from packaging import version

from requests.exceptions import HTTPError

from tqdm.auto import tqdm

import diffusers

from .. import __version__

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,

DEPRECATED_REVISION_ARGS,

DIFFUSERS_CACHE,

HF_HUB_OFFLINE,

SAFETENSORS_WEIGHTS_NAME,

WEIGHTS_NAME,

BaseOutput,

deprecate,

get_class_from_dynamic_module,

is_accelerate_available,

is_accelerate_version,

is_compiled_module,

is_torch_version,

is_transformers_available,

logging,

numpy_to_pil,

)

if is_transformers_available():

import transformers

from transformers import PreTrainedModel

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

from ..utils import FLAX_WEIGHTS_NAME, ONNX_EXTERNAL_WEIGHTS_NAME, ONNX_WEIGHTS_NAME, PushToHubMixin

if is_accelerate_available():

import accelerate

INDEX_FILE = "diffusion_pytorch_model.bin"

CUSTOM_PIPELINE_FILE_NAME = "pipeline.py"

DUMMY_MODULES_FOLDER = "diffusers.utils"

TRANSFORMERS_DUMMY_MODULES_FOLDER = "transformers.utils"

CONNECTED_PIPES_KEYS = ["prior"]

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`)

List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,

num_channels)`.

"""

images: Union[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

def is_safetensors_compatible(filenames, variant=None, passed_components=None) -> bool:

"""

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()

passed_components = passed_components or []

for filename in filenames:

_, extension = os.path.splitext(filename)

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

continue

if extension == ".bin":

pt_filenames.append(filename)

elif extension == ".safetensors":

sf_filenames.add(filename)

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)

if filename.startswith("pytorch_model"):

filename = filename.replace("pytorch_model", "model")

else:

filename = filename

expected_sf_filename = os.path.join(path, filename)

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

def variant_compatible_siblings(filenames, variant=None) -> Union[List[os.PathLike], str]:

weight_names = [

WEIGHTS_NAME,

SAFETENSORS_WEIGHTS_NAME,

FLAX_WEIGHTS_NAME,

ONNX_WEIGHTS_NAME,

ONNX_EXTERNAL_WEIGHTS_NAME,

]

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]

# -00001-of-00002

transformers_index_format = r"\d{5}-of-\d{5}"

if variant is not None:

# `diffusion_pytorch_model.fp16.bin` as well as `model.fp16-00001-of-00002.safetensors`

variant_file_re = re.compile(

rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$"

)

# `text_encoder/pytorch_model.bin.index.fp16.json`

variant_index_re = re.compile(

rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.{variant}\.json$"

)

# `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors`

non_variant_file_re = re.compile(

rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$"

)

# `text_encoder/pytorch_model.bin.index.json`

non_variant_index_re = re.compile(rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.json")

if variant is not None:

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

else:

variant_filenames = set()

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

# all variant filenames will be used by default

usable_filenames = set(variant_filenames)

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

for f in non_variant_filenames:

variant_filename = convert_to_variant(f)

if variant_filename not in usable_filenames:

usable_filenames.add(f)

return usable_filenames, variant_filenames

def warn_deprecated_model_variant(pretrained_model_name_or_path, use_auth_token, variant, revision, model_filenames):

info = model_info(

pretrained_model_name_or_path,

use_auth_token=use_auth_token,

revision=None,

)

filenames = {sibling.rfilename for sibling in info.siblings}

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]

if set(comp_model_filenames) == set(model_filenames):

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,

)

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

# 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]

model_cls = sub_model.__class__

if is_compiled_module(sub_model):

model_cls = sub_model._orig_mod.__class__

if not issubclass(model_cls, expected_class_obj):

raise ValueError(

f"{passed_class_obj[name]} is of type: {model_cls}, but should be" f" {expected_class_obj}"

)

else:

logger.warning(

f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it"

" has the correct type"

)

def get_class_obj_and_candidates(library_name, class_name, importable_classes, pipelines, is_pipeline_module):

"""Simple helper method to retrieve class object of module as well as potential parent class objects"""

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()}

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

def _get_pipeline_class(

class_obj, config, load_connected_pipeline=False, custom_pipeline=None, cache_dir=None, revision=None

):

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()

else:

file_name = CUSTOM_PIPELINE_FILE_NAME

return get_class_from_dynamic_module(

custom_pipeline, module_file=file_name, cache_dir=cache_dir, revision=revision

)

if class_obj != DiffusionPipeline:

return class_obj

diffusers_module = importlib.import_module(class_obj.__module__.split(".")[0])

pipeline_cls = getattr(diffusers_module, config["_class_name"])

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

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]],

max_memory: Optional[Dict[Union[int, str], Union[int, str]]],

offload_folder: Optional[Union[str, os.PathLike]],

offload_state_dict: bool,

model_variants: Dict[str, str],

name: str,

from_flax: bool,

variant: str,

low_cpu_mem_usage: bool,

cached_folder: Union[str, os.PathLike],

):

"""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(

library_name, class_name, importable_classes, pipelines, is_pipeline_module

)

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

loading_kwargs = {}

if issubclass(class_obj, torch.nn.Module):

loading_kwargs["torch_dtype"] = torch_dtype

if issubclass(class_obj, diffusers.OnnxRuntimeModel):

loading_kwargs["provider"] = provider

loading_kwargs["sess_options"] = sess_options

is_diffusers_model = issubclass(class_obj, diffusers.ModelMixin)

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

loading_kwargs["max_memory"] = max_memory

loading_kwargs["offload_folder"] = offload_folder

loading_kwargs["offload_state_dict"] = offload_state_dict

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

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"

_optional_components = []

_exclude_from_cpu_offload = []

_load_connected_pipes = False

_is_onnx = False

def register_modules(self, **kwargs):

# import it here to avoid circular import

from diffusers import pipelines

for name, module in kwargs.items():

# retrieve library

if module is None:

register_dict = {name: (None, None)}

else:

# register the config from the original module, not the dynamo compiled one

if is_compiled_module(module):

not_compiled_module = module._orig_mod

else:

not_compiled_module = module

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

# check if the module is a pipeline module

module_path_items = not_compiled_module.__module__.split(".")

pipeline_dir = module_path_items[-2] if len(module_path_items) > 2 else None

path = not_compiled_module.__module__.split(".")

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.

if is_pipeline_module:

library = pipeline_dir

elif library not in LOADABLE_CLASSES:

library = not_compiled_module.__module__

# retrieve class_name

class_name = not_compiled_module.__class__.__name__

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 = (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)

def save_pretrained(

self,

save_directory: Union[str, os.PathLike],

safe_serialization: bool = True,

variant: Optional[str] = None,

push_to_hub: 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`.

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", 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

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 = sub_model._orig_mod

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.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

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_kwargs = {}

if save_method_accept_safe:

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)

# finally save the config

self.save_config(save_directory)

if push_to_hub:

self._upload_folder(

save_directory,

repo_id,

token=token,

commit_message=commit_message,

create_pr=create_pr,

)

def to(

self,

torch_device: Optional[Union[str, torch.device]] = None,

torch_dtype: Optional[torch.dtype] = None,

silence_dtype_warnings: bool = False,

):

if torch_device is None and torch_dtype is None:

return self

# 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

return hasattr(module, "_hf_hook") and not isinstance(

module._hf_hook, (accelerate.hooks.CpuOffload, 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()

)

if pipeline_is_sequentially_offloaded and torch_device and torch.device(torch_device).type == "cuda":

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())

if pipeline_is_offloaded and torch_device and torch.device(torch_device).type == "cuda":

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."

)

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_8bit = hasattr(module, "is_loaded_in_8bit") and module.is_loaded_in_8bit

if is_loaded_in_8bit and torch_dtype is not None:

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."

)

if is_loaded_in_8bit and torch_device is not None:

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:

module.to(torch_device, torch_dtype)

if (

module.dtype == torch.float16

and str(torch_device) in ["cpu"]

and not silence_dtype_warnings

and not is_offloaded

):

logger.warning(

"Pipelines loaded with `torch_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")

@classmethod

def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):

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 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.

```

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`].

torch_dtype (`str` or `torch.dtype`, *optional*):

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.

custom_pipeline (`str`, *optional*):

<Tip warning={true}>

🧪 This is an experimental feature and may change in the future.

</Tip>

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.

resume_download (`bool`, *optional*, defaults to `False`):

Whether or not to resume downloading the model weights and configuration files. If set to `False`, any

incompletely downloaded files are deleted.

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.

use_auth_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*, defaults to `"main"`):

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. It can be a 🤗 Diffusers version when loading a

custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.

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` or `Dict[str, Union[int, str, torch.device]]`, *optional*):

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

same device.

Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For

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).

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`.

<Tip>

To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with

`huggingface-cli login`.

</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

```

"""

cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)

resume_download = kwargs.pop("resume_download", False)

force_download = kwargs.pop("force_download", False)

proxies = kwargs.pop("proxies", None)

local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)

use_auth_token = kwargs.pop("use_auth_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)

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", False)

low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)

variant = kwargs.pop("variant", None)

use_safetensors = kwargs.pop("use_safetensors", None)

use_onnx = kwargs.pop("use_onnx", None)

load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)

# 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):

cached_folder = cls.download(

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,

use_auth_token=use_auth_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,

load_connected_pipeline=load_connected_pipeline,

**kwargs,

)

else:

cached_folder = pretrained_model_name_or_path

config_dict = cls.load_config(cached_folder)

# 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

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

variant_exists = is_folder and any(

p.split(".")[1].startswith(variant) for p in os.listdir(folder_path)

)

if variant_exists:

model_variants[folder] = variant

# 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

pipeline_class = _get_pipeline_class(

cls,

config_dict,

load_connected_pipeline=load_connected_pipeline,

custom_pipeline=custom_pipeline,

cache_dir=cache_dir,

revision=custom_revision,

)

# DEPRECATED: To be removed in 1.0.0

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."

)