# coding=utf-8

# Copyright 2024 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

from dataclasses import dataclass

from pathlib import Path

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

import numpy as np

import PIL.Image

import requests

import torch

from huggingface_hub import (

ModelCard,

create_repo,

hf_hub_download,

model_info,

snapshot_download,

)

from huggingface_hub.utils import OfflineModeIsEnabled, validate_hf_hub_args

from packaging import version

from requests.exceptions import HTTPError

from tqdm.auto import tqdm

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 ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME

from ..utils import (

CONFIG_NAME,

DEPRECATED_REVISION_ARGS,

BaseOutput,

PushToHubMixin,

deprecate,

is_accelerate_available,

is_accelerate_version,

is_torch_npu_available,

is_torch_version,

logging,

numpy_to_pil,

)

from ..utils.hub_utils import load_or_create_model_card, populate_model_card

from ..utils.torch_utils import 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,

_fetch_class_library_tuple,

_get_custom_pipeline_class,

_get_final_device_map,

_get_pipeline_class,

_unwrap_model,

is_safetensors_compatible,

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"]

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: 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

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: 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 = _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_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:

# 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):

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`.

"""

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

# 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 (

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()

)

if pipeline_is_sequentially_offloaded and device and torch.device(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."

)

is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1

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()` first and then call `to()`."

)

# 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 and torch.device(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 dtype is not None:

logger.warning(

f"The module '{module.__class__.__name__}' has been loaded in 8bit and conversion to {dtype} is not yet supported. Module is still in 8bit precision."

)

if is_loaded_in_8bit and device is not None:

logger.warning(

f"The module '{module.__class__.__name__}' has been loaded in 8bit and moving it to {dtype} via `.to()` is not yet supported. Module is still on {module.device}."

)

else:

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: 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:

Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1

of Diffusers.

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` 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", None)

resume_download = kwargs.pop("resume_download", 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)

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)

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 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`."

)

# 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,

resume_download=resume_download,

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,

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

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]

pipeline_class = _get_pipeline_class(

cls,

config_dict,

load_connected_pipeline=load_connected_pipeline,

custom_pipeline=custom_pipeline,

class_name=custom_class_name,

cache_dir=cache_dir,

revision=custom_revision,

)

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

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)

# 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)

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,

passed_class_obj=passed_class_obj,

init_dict=init_dict,

library=library,

max_memory=max_memory,

torch_dtype=torch_dtype,

cached_folder=cached_folder,

force_download=force_download,

resume_download=resume_download,

proxies=proxies,

local_files_only=local_files_only,

token=token,

revision=revision,

)

# 7. Load each module in the pipeline

current_device_map = None

for name, (library_name, class_name) in logging.tqdm(init_dict.items(), desc="Loading pipeline components..."):

if final_device_map is not None and len(final_device_map) > 0:

component_device = final_device_map.get(name, None)

if component_device is not None:

current_device_map = {"": component_device}

else:

current_device_map = None

# 7.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names

class_name = class_name[4:] if class_name.startswith("Flax") else class_name

# 7.2 Define all importable classes

is_pipeline_module = hasattr(pipelines, library_name)

importable_classes = ALL_IMPORTABLE_CLASSES

loaded_sub_model = None

# 7.3 Use passed sub model or load class_name from library_name

if name in passed_class_obj:

# 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

)

loaded_sub_model = passed_class_obj[name]

else:

# 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=current_device_map,

max_memory=max_memory,

offload_folder=offload_folder,

offload_state_dict=offload_state_dict,

model_variants=model_variants,

name=name,

from_flax=from_flax,

variant=variant,

low_cpu_mem_usage=low_cpu_mem_usage,

cached_folder=cached_folder,

)

logger.info(

f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}."

)

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

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,

"token": token,

"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,

}

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

connected_pipes = {

prefix: DiffusionPipeline.from_pretrained(

repo_id, **load_kwargs.copy(), **get_connected_passed_kwargs(prefix)

)

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()}

)

# 8. Potentially add passed objects if expected

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."

)

# 10. Instantiate the pipeline

model = pipeline_class(**init_kwargs)

# 11. Save where the model was instantiated from

model.register_to_config(_name_or_path=pretrained_model_name_or_path)

if device_map is not None:

setattr(model, "hf_device_map", final_device_map)

return model

@property

def name_or_path(self) -> str:

return getattr(self.config, "_name_or_path", None)

@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")