[Refactor] refactor `loaders.py` to make it cleaner and leaner. (#5771)

* refactor loaders.py to make it cleaner and leaner.

* refactor loaders init

* inits.

* textual inversion to the init.

* inits.

* remove certain modules from the main init.

* AttnProcsLayers

* fix imports

* avoid circular import.

* fix circular import pt 2.

* address PR comments

* imports

* fix: imports.

* remove from main init for avoiding circular deps.

* remove spurious deps.

* fix-copies.

* fix imports.

* more debug

* more debug

* Apply suggestions from code review

* Apply suggestions from code review

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

examples/dreambooth/train_dreambooth_lora.py

@@ -51,16 +51,13 @@ from diffusers import (
     StableDiffusionPipeline,
     UNet2DConditionModel,
 )
-from diffusers.loaders import (
-    LoraLoaderMixin,
-    text_encoder_lora_state_dict,
-)
+from diffusers.loaders import LoraLoaderMixin
 from diffusers.models.attention_processor import (
     AttnAddedKVProcessor,
     AttnAddedKVProcessor2_0,
     SlicedAttnAddedKVProcessor,
 )
-from diffusers.models.lora import LoRALinearLayer
+from diffusers.models.lora import LoRALinearLayer, text_encoder_lora_state_dict
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import unet_lora_state_dict
 from diffusers.utils import check_min_version, is_wandb_available

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -49,8 +49,8 @@ from diffusers import (
     StableDiffusionXLPipeline,
     UNet2DConditionModel,
 )
-from diffusers.loaders import LoraLoaderMixin, text_encoder_lora_state_dict
-from diffusers.models.lora import LoRALinearLayer
+from diffusers.loaders import LoraLoaderMixin
+from diffusers.models.lora import LoRALinearLayer, text_encoder_lora_state_dict
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import unet_lora_state_dict
 from diffusers.utils import check_min_version, is_wandb_available

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -49,8 +49,8 @@ from diffusers import (
     StableDiffusionXLPipeline,
     UNet2DConditionModel,
 )
-from diffusers.loaders import LoraLoaderMixin, text_encoder_lora_state_dict
-from diffusers.models.lora import LoRALinearLayer
+from diffusers.loaders import LoraLoaderMixin
+from diffusers.models.lora import LoRALinearLayer, text_encoder_lora_state_dict
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_snr
 from diffusers.utils import check_min_version, is_wandb_available

src/diffusers/__init__.py

@@ -94,6 +94,7 @@ else:
             "VQModel",
         ]
     )
+
     _import_structure["optimization"] = [
         "get_constant_schedule",
         "get_constant_schedule_with_warmup",
@@ -103,7 +104,6 @@ else:
         "get_polynomial_decay_schedule_with_warmup",
         "get_scheduler",
     ]
-
     _import_structure["pipelines"].extend(
         [
             "AudioPipelineOutput",

src/diffusers/loaders/__init__.py

+from typing import TYPE_CHECKING
+
+from ..utils import DIFFUSERS_SLOW_IMPORT, _LazyModule, deprecate
+from ..utils.import_utils import is_torch_available, is_transformers_available
+
+
+def text_encoder_lora_state_dict(text_encoder):
+    deprecate(
+        "text_encoder_load_state_dict in `models`",
+        "0.27.0",
+        "`text_encoder_lora_state_dict` has been moved to `diffusers.models.lora`. Please make sure to import it via `from diffusers.models.lora import text_encoder_lora_state_dict`.",
+    )
+    state_dict = {}
+
+    for name, module in text_encoder_attn_modules(text_encoder):
+        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
+
+    return state_dict
+
+
+if is_transformers_available():
+
+    def text_encoder_attn_modules(text_encoder):
+        deprecate(
+            "text_encoder_attn_modules in `models`",
+            "0.27.0",
+            "`text_encoder_lora_state_dict` has been moved to `diffusers.models.lora`. Please make sure to import it via `from diffusers.models.lora import text_encoder_lora_state_dict`.",
+        )
+        from transformers import CLIPTextModel, CLIPTextModelWithProjection
+
+        attn_modules = []
+
+        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+                name = f"text_model.encoder.layers.{i}.self_attn"
+                mod = layer.self_attn
+                attn_modules.append((name, mod))
+        else:
+            raise ValueError(f"do not know how to get attention modules for: {text_encoder.__class__.__name__}")
+
+        return attn_modules
+
+
+_import_structure = {}
+
+if is_torch_available():
+    _import_structure["single_file"] = ["FromOriginalControlnetMixin", "FromOriginalVAEMixin"]
+    _import_structure["unet"] = ["UNet2DConditionLoadersMixin"]
+    _import_structure["utils"] = ["AttnProcsLayers"]
+
+    if is_transformers_available():
+        _import_structure["single_file"].extend(["FromSingleFileMixin"])
+        _import_structure["lora"] = ["LoraLoaderMixin", "StableDiffusionXLLoraLoaderMixin"]
+        _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    if is_torch_available():
+        from ..models.lora import text_encoder_lora_state_dict
+        from .single_file import FromOriginalControlnetMixin, FromOriginalVAEMixin
+        from .unet import UNet2DConditionLoadersMixin
+        from .utils import AttnProcsLayers
+
+        if is_transformers_available():
+            from .lora import LoraLoaderMixin, StableDiffusionXLLoraLoaderMixin
+            from .single_file import FromSingleFileMixin
+            from .textual_inversion import TextualInversionLoaderMixin
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/diffusers/loaders.py → src/diffusers/loaders/lora.py

@@ -13,22 +13,18 @@
 # limitations under the License.
 import os
 import re
-from collections import defaultdict
 from contextlib import nullcontext
-from io import BytesIO
-from pathlib import Path
 from typing import Callable, Dict, List, Optional, Union
 
-import requests
 import safetensors
 import torch
-from huggingface_hub import hf_hub_download, model_info
+from huggingface_hub import model_info
 from packaging import version
 from torch import nn
 
-from . import __version__
-from .models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
-from .utils import (
+from .. import __version__
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+from ..utils import (
     DIFFUSERS_CACHE,
     HF_HUB_OFFLINE,
     USE_PEFT_BACKEND,
@@ -41,7 +37,6 @@ from .utils import (
     get_adapter_name,
     get_peft_kwargs,
     is_accelerate_available,
-    is_omegaconf_available,
     is_transformers_available,
     logging,
     recurse_remove_peft_layers,
@@ -49,1138 +44,26 @@ from .utils import (
     set_adapter_layers,
     set_weights_and_activate_adapters,
 )
-from .utils.import_utils import BACKENDS_MAPPING
 
 
 if is_transformers_available():
-    from transformers import CLIPTextModel, CLIPTextModelWithProjection, PreTrainedModel
+    from transformers import PreTrainedModel
 
-if is_accelerate_available():
-    from accelerate import init_empty_weights
-    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
-
-logger = logging.get_logger(__name__)
-
-TEXT_ENCODER_NAME = "text_encoder"
-UNET_NAME = "unet"
-
-LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
-LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
-
-TEXT_INVERSION_NAME = "learned_embeds.bin"
-TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
-
-CUSTOM_DIFFUSION_WEIGHT_NAME = "pytorch_custom_diffusion_weights.bin"
-CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE = "pytorch_custom_diffusion_weights.safetensors"
-
-LORA_DEPRECATION_MESSAGE = "You are using an old version of LoRA backend. This will be deprecated in the next releases in favor of PEFT make sure to install the latest PEFT and transformers packages in the future."
-
-
-class PatchedLoraProjection(nn.Module):
-    def __init__(self, regular_linear_layer, lora_scale=1, network_alpha=None, rank=4, dtype=None):
-        super().__init__()
-        from .models.lora import LoRALinearLayer
-
-        self.regular_linear_layer = regular_linear_layer
-
-        device = self.regular_linear_layer.weight.device
-
-        if dtype is None:
-            dtype = self.regular_linear_layer.weight.dtype
-
-        self.lora_linear_layer = LoRALinearLayer(
-            self.regular_linear_layer.in_features,
-            self.regular_linear_layer.out_features,
-            network_alpha=network_alpha,
-            device=device,
-            dtype=dtype,
-            rank=rank,
-        )
-
-        self.lora_scale = lora_scale
-
-    # overwrite PyTorch's `state_dict` to be sure that only the 'regular_linear_layer' weights are saved
-    # when saving the whole text encoder model and when LoRA is unloaded or fused
-    def state_dict(self, *args, destination=None, prefix="", keep_vars=False):
-        if self.lora_linear_layer is None:
-            return self.regular_linear_layer.state_dict(
-                *args, destination=destination, prefix=prefix, keep_vars=keep_vars
-            )
-
-        return super().state_dict(*args, destination=destination, prefix=prefix, keep_vars=keep_vars)
-
-    def _fuse_lora(self, lora_scale=1.0, safe_fusing=False):
-        if self.lora_linear_layer is None:
-            return
-
-        dtype, device = self.regular_linear_layer.weight.data.dtype, self.regular_linear_layer.weight.data.device
-
-        w_orig = self.regular_linear_layer.weight.data.float()
-        w_up = self.lora_linear_layer.up.weight.data.float()
-        w_down = self.lora_linear_layer.down.weight.data.float()
-
-        if self.lora_linear_layer.network_alpha is not None:
-            w_up = w_up * self.lora_linear_layer.network_alpha / self.lora_linear_layer.rank
-
-        fused_weight = w_orig + (lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
-
-        if safe_fusing and torch.isnan(fused_weight).any().item():
-            raise ValueError(
-                "This LoRA weight seems to be broken. "
-                f"Encountered NaN values when trying to fuse LoRA weights for {self}."
-                "LoRA weights will not be fused."
-            )
-
-        self.regular_linear_layer.weight.data = fused_weight.to(device=device, dtype=dtype)
-
-        # we can drop the lora layer now
-        self.lora_linear_layer = None
-
-        # offload the up and down matrices to CPU to not blow the memory
-        self.w_up = w_up.cpu()
-        self.w_down = w_down.cpu()
-        self.lora_scale = lora_scale
-
-    def _unfuse_lora(self):
-        if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
-            return
-
-        fused_weight = self.regular_linear_layer.weight.data
-        dtype, device = fused_weight.dtype, fused_weight.device
-
-        w_up = self.w_up.to(device=device).float()
-        w_down = self.w_down.to(device).float()
-
-        unfused_weight = fused_weight.float() - (self.lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
-        self.regular_linear_layer.weight.data = unfused_weight.to(device=device, dtype=dtype)
-
-        self.w_up = None
-        self.w_down = None
-
-    def forward(self, input):
-        if self.lora_scale is None:
-            self.lora_scale = 1.0
-        if self.lora_linear_layer is None:
-            return self.regular_linear_layer(input)
-        return self.regular_linear_layer(input) + (self.lora_scale * self.lora_linear_layer(input))
-
-
-def text_encoder_attn_modules(text_encoder):
-    attn_modules = []
-
-    if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-        for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-            name = f"text_model.encoder.layers.{i}.self_attn"
-            mod = layer.self_attn
-            attn_modules.append((name, mod))
-    else:
-        raise ValueError(f"do not know how to get attention modules for: {text_encoder.__class__.__name__}")
-
-    return attn_modules
-
-
-def text_encoder_mlp_modules(text_encoder):
-    mlp_modules = []
-
-    if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-        for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-            mlp_mod = layer.mlp
-            name = f"text_model.encoder.layers.{i}.mlp"
-            mlp_modules.append((name, mlp_mod))
-    else:
-        raise ValueError(f"do not know how to get mlp modules for: {text_encoder.__class__.__name__}")
-
-    return mlp_modules
-
-
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
-class AttnProcsLayers(torch.nn.Module):
-    def __init__(self, state_dict: Dict[str, torch.Tensor]):
-        super().__init__()
-        self.layers = torch.nn.ModuleList(state_dict.values())
-        self.mapping = dict(enumerate(state_dict.keys()))
-        self.rev_mapping = {v: k for k, v in enumerate(state_dict.keys())}
-
-        # .processor for unet, .self_attn for text encoder
-        self.split_keys = [".processor", ".self_attn"]
-
-        # we add a hook to state_dict() and load_state_dict() so that the
-        # naming fits with `unet.attn_processors`
-        def map_to(module, state_dict, *args, **kwargs):
-            new_state_dict = {}
-            for key, value in state_dict.items():
-                num = int(key.split(".")[1])  # 0 is always "layers"
-                new_key = key.replace(f"layers.{num}", module.mapping[num])
-                new_state_dict[new_key] = value
-
-            return new_state_dict
-
-        def remap_key(key, state_dict):
-            for k in self.split_keys:
-                if k in key:
-                    return key.split(k)[0] + k
-
-            raise ValueError(
-                f"There seems to be a problem with the state_dict: {set(state_dict.keys())}. {key} has to have one of {self.split_keys}."
-            )
-
-        def map_from(module, state_dict, *args, **kwargs):
-            all_keys = list(state_dict.keys())
-            for key in all_keys:
-                replace_key = remap_key(key, state_dict)
-                new_key = key.replace(replace_key, f"layers.{module.rev_mapping[replace_key]}")
-                state_dict[new_key] = state_dict[key]
-                del state_dict[key]
-
-        self._register_state_dict_hook(map_to)
-        self._register_load_state_dict_pre_hook(map_from, with_module=True)
-
-
-class UNet2DConditionLoadersMixin:
-    text_encoder_name = TEXT_ENCODER_NAME
-    unet_name = UNET_NAME
-
-    def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
-        r"""
-        Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
-        defined in
-        [`attention_processor.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py)
-        and be a `torch.nn.Module` class.
-
-        Parameters:
-            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
-                Can be either:
-
-                    - A string, the model id (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
-                      the Hub.
-                    - A path to a directory (for example `./my_model_directory`) containing the model weights saved
-                      with [`ModelMixin.save_pretrained`].
-                    - A [torch state
-                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
-
-            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.
-            force_download (`bool`, *optional*, defaults to `False`):
-                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
-                cached versions if they exist.
-            resume_download (`bool`, *optional*, defaults to `False`):
-                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.
-            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.
-            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.
-            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.
-            subfolder (`str`, *optional*, defaults to `""`):
-                The subfolder location of a model file within a larger model repository on the Hub or locally.
-            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.
-
-        """
-        from .models.attention_processor import (
-            CustomDiffusionAttnProcessor,
-        )
-        from .models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
-
-        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
-        force_download = kwargs.pop("force_download", False)
-        resume_download = kwargs.pop("resume_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)
-        subfolder = kwargs.pop("subfolder", None)
-        weight_name = kwargs.pop("weight_name", None)
-        use_safetensors = kwargs.pop("use_safetensors", None)
-        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
-        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
-        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
-        network_alphas = kwargs.pop("network_alphas", None)
-
-        _pipeline = kwargs.pop("_pipeline", None)
-
-        is_network_alphas_none = network_alphas is None
-
-        allow_pickle = False
-
-        if use_safetensors is None:
-            use_safetensors = True
-            allow_pickle = True
-
-        user_agent = {
-            "file_type": "attn_procs_weights",
-            "framework": "pytorch",
-        }
-
-        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."
-            )
-
-        model_file = None
-        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
-            # Let's first try to load .safetensors weights
-            if (use_safetensors and weight_name is None) or (
-                weight_name is not None and weight_name.endswith(".safetensors")
-            ):
-                try:
-                    model_file = _get_model_file(
-                        pretrained_model_name_or_path_or_dict,
-                        weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
-                        cache_dir=cache_dir,
-                        force_download=force_download,
-                        resume_download=resume_download,
-                        proxies=proxies,
-                        local_files_only=local_files_only,
-                        use_auth_token=use_auth_token,
-                        revision=revision,
-                        subfolder=subfolder,
-                        user_agent=user_agent,
-                    )
-                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
-                except IOError as e:
-                    if not allow_pickle:
-                        raise e
-                    # try loading non-safetensors weights
-                    pass
-            if model_file is None:
-                model_file = _get_model_file(
-                    pretrained_model_name_or_path_or_dict,
-                    weights_name=weight_name or LORA_WEIGHT_NAME,
-                    cache_dir=cache_dir,
-                    force_download=force_download,
-                    resume_download=resume_download,
-                    proxies=proxies,
-                    local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
-                    revision=revision,
-                    subfolder=subfolder,
-                    user_agent=user_agent,
-                )
-                state_dict = torch.load(model_file, map_location="cpu")
-        else:
-            state_dict = pretrained_model_name_or_path_or_dict
-
-        # fill attn processors
-        lora_layers_list = []
-
-        is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys()) and not USE_PEFT_BACKEND
-        is_custom_diffusion = any("custom_diffusion" in k for k in state_dict.keys())
-
-        if is_lora:
-            # correct keys
-            state_dict, network_alphas = self.convert_state_dict_legacy_attn_format(state_dict, network_alphas)
-
-            if network_alphas is not None:
-                network_alphas_keys = list(network_alphas.keys())
-                used_network_alphas_keys = set()
-
-            lora_grouped_dict = defaultdict(dict)
-            mapped_network_alphas = {}
-
-            all_keys = list(state_dict.keys())
-            for key in all_keys:
-                value = state_dict.pop(key)
-                attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
-                lora_grouped_dict[attn_processor_key][sub_key] = value
-
-                # Create another `mapped_network_alphas` dictionary so that we can properly map them.
-                if network_alphas is not None:
-                    for k in network_alphas_keys:
-                        if k.replace(".alpha", "") in key:
-                            mapped_network_alphas.update({attn_processor_key: network_alphas.get(k)})
-                            used_network_alphas_keys.add(k)
-
-            if not is_network_alphas_none:
-                if len(set(network_alphas_keys) - used_network_alphas_keys) > 0:
-                    raise ValueError(
-                        f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
-                    )
-
-            if len(state_dict) > 0:
-                raise ValueError(
-                    f"The `state_dict` has to be empty at this point but has the following keys \n\n {', '.join(state_dict.keys())}"
-                )
-
-            for key, value_dict in lora_grouped_dict.items():
-                attn_processor = self
-                for sub_key in key.split("."):
-                    attn_processor = getattr(attn_processor, sub_key)
-
-                # Process non-attention layers, which don't have to_{k,v,q,out_proj}_lora layers
-                # or add_{k,v,q,out_proj}_proj_lora layers.
-                rank = value_dict["lora.down.weight"].shape[0]
-
-                if isinstance(attn_processor, LoRACompatibleConv):
-                    in_features = attn_processor.in_channels
-                    out_features = attn_processor.out_channels
-                    kernel_size = attn_processor.kernel_size
-
-                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
-                    with ctx():
-                        lora = LoRAConv2dLayer(
-                            in_features=in_features,
-                            out_features=out_features,
-                            rank=rank,
-                            kernel_size=kernel_size,
-                            stride=attn_processor.stride,
-                            padding=attn_processor.padding,
-                            network_alpha=mapped_network_alphas.get(key),
-                        )
-                elif isinstance(attn_processor, LoRACompatibleLinear):
-                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
-                    with ctx():
-                        lora = LoRALinearLayer(
-                            attn_processor.in_features,
-                            attn_processor.out_features,
-                            rank,
-                            mapped_network_alphas.get(key),
-                        )
-                else:
-                    raise ValueError(f"Module {key} is not a LoRACompatibleConv or LoRACompatibleLinear module.")
-
-                value_dict = {k.replace("lora.", ""): v for k, v in value_dict.items()}
-                lora_layers_list.append((attn_processor, lora))
-
-                if low_cpu_mem_usage:
-                    device = next(iter(value_dict.values())).device
-                    dtype = next(iter(value_dict.values())).dtype
-                    load_model_dict_into_meta(lora, value_dict, device=device, dtype=dtype)
-                else:
-                    lora.load_state_dict(value_dict)
-
-        elif is_custom_diffusion:
-            attn_processors = {}
-            custom_diffusion_grouped_dict = defaultdict(dict)
-            for key, value in state_dict.items():
-                if len(value) == 0:
-                    custom_diffusion_grouped_dict[key] = {}
-                else:
-                    if "to_out" in key:
-                        attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
-                    else:
-                        attn_processor_key, sub_key = ".".join(key.split(".")[:-2]), ".".join(key.split(".")[-2:])
-                    custom_diffusion_grouped_dict[attn_processor_key][sub_key] = value
-
-            for key, value_dict in custom_diffusion_grouped_dict.items():
-                if len(value_dict) == 0:
-                    attn_processors[key] = CustomDiffusionAttnProcessor(
-                        train_kv=False, train_q_out=False, hidden_size=None, cross_attention_dim=None
-                    )
-                else:
-                    cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[1]
-                    hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[0]
-                    train_q_out = True if "to_q_custom_diffusion.weight" in value_dict else False
-                    attn_processors[key] = CustomDiffusionAttnProcessor(
-                        train_kv=True,
-                        train_q_out=train_q_out,
-                        hidden_size=hidden_size,
-                        cross_attention_dim=cross_attention_dim,
-                    )
-                    attn_processors[key].load_state_dict(value_dict)
-        elif USE_PEFT_BACKEND:
-            # In that case we have nothing to do as loading the adapter weights is already handled above by `set_peft_model_state_dict`
-            # on the Unet
-            pass
-        else:
-            raise ValueError(
-                f"{model_file} does not seem to be in the correct format expected by LoRA or Custom Diffusion training."
-            )
-
-        # <Unsafe code
-        # We can be sure that the following works as it just sets attention processors, lora layers and puts all in the same dtype
-        # Now we remove any existing hooks to
-        is_model_cpu_offload = False
-        is_sequential_cpu_offload = False
-
-        # For PEFT backend the Unet is already offloaded at this stage as it is handled inside `lora_lora_weights_into_unet`
-        if not USE_PEFT_BACKEND:
-            if _pipeline is not None:
-                for _, component in _pipeline.components.items():
-                    if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
-                        is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
-                        is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
-
-                        logger.info(
-                            "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
-                        )
-                        remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
-
-            # only custom diffusion needs to set attn processors
-            if is_custom_diffusion:
-                self.set_attn_processor(attn_processors)
-
-            # set lora layers
-            for target_module, lora_layer in lora_layers_list:
-                target_module.set_lora_layer(lora_layer)
-
-            self.to(dtype=self.dtype, device=self.device)
-
-            # Offload back.
-            if is_model_cpu_offload:
-                _pipeline.enable_model_cpu_offload()
-            elif is_sequential_cpu_offload:
-                _pipeline.enable_sequential_cpu_offload()
-            # Unsafe code />
-
-    def convert_state_dict_legacy_attn_format(self, state_dict, network_alphas):
-        is_new_lora_format = all(
-            key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
-        )
-        if is_new_lora_format:
-            # Strip the `"unet"` prefix.
-            is_text_encoder_present = any(key.startswith(self.text_encoder_name) for key in state_dict.keys())
-            if is_text_encoder_present:
-                warn_message = "The state_dict contains LoRA params corresponding to the text encoder which are not being used here. To use both UNet and text encoder related LoRA params, use [`pipe.load_lora_weights()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights)."
-                logger.warn(warn_message)
-            unet_keys = [k for k in state_dict.keys() if k.startswith(self.unet_name)]
-            state_dict = {k.replace(f"{self.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
-
-        # change processor format to 'pure' LoRACompatibleLinear format
-        if any("processor" in k.split(".") for k in state_dict.keys()):
-
-            def format_to_lora_compatible(key):
-                if "processor" not in key.split("."):
-                    return key
-                return key.replace(".processor", "").replace("to_out_lora", "to_out.0.lora").replace("_lora", ".lora")
-
-            state_dict = {format_to_lora_compatible(k): v for k, v in state_dict.items()}
-
-            if network_alphas is not None:
-                network_alphas = {format_to_lora_compatible(k): v for k, v in network_alphas.items()}
-        return state_dict, network_alphas
-
-    def save_attn_procs(
-        self,
-        save_directory: Union[str, os.PathLike],
-        is_main_process: bool = True,
-        weight_name: str = None,
-        save_function: Callable = None,
-        safe_serialization: bool = True,
-        **kwargs,
-    ):
-        r"""
-        Save an attention processor to a directory so that it can be reloaded using the
-        [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] method.
-
-        Arguments:
-            save_directory (`str` or `os.PathLike`):
-                Directory to save an attention processor to. Will be created if it doesn't exist.
-            is_main_process (`bool`, *optional*, defaults to `True`):
-                Whether the process calling this is the main process or not. Useful during distributed training and you
-                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
-                process to avoid race conditions.
-            save_function (`Callable`):
-                The function to use to save the state dictionary. Useful during distributed training when you need to
-                replace `torch.save` with another method. Can be configured with the environment variable
-                `DIFFUSERS_SAVE_MODE`.
-            safe_serialization (`bool`, *optional*, defaults to `True`):
-                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
-        """
-        from .models.attention_processor import (
-            CustomDiffusionAttnProcessor,
-            CustomDiffusionAttnProcessor2_0,
-            CustomDiffusionXFormersAttnProcessor,
-        )
-
-        if os.path.isfile(save_directory):
-            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
-            return
-
-        if save_function is None:
-            if safe_serialization:
-
-                def save_function(weights, filename):
-                    return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
-
-            else:
-                save_function = torch.save
-
-        os.makedirs(save_directory, exist_ok=True)
-
-        is_custom_diffusion = any(
-            isinstance(
-                x,
-                (CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor),
-            )
-            for (_, x) in self.attn_processors.items()
-        )
-        if is_custom_diffusion:
-            model_to_save = AttnProcsLayers(
-                {
-                    y: x
-                    for (y, x) in self.attn_processors.items()
-                    if isinstance(
-                        x,
-                        (
-                            CustomDiffusionAttnProcessor,
-                            CustomDiffusionAttnProcessor2_0,
-                            CustomDiffusionXFormersAttnProcessor,
-                        ),
-                    )
-                }
-            )
-            state_dict = model_to_save.state_dict()
-            for name, attn in self.attn_processors.items():
-                if len(attn.state_dict()) == 0:
-                    state_dict[name] = {}
-        else:
-            model_to_save = AttnProcsLayers(self.attn_processors)
-            state_dict = model_to_save.state_dict()
-
-        if weight_name is None:
-            if safe_serialization:
-                weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE if is_custom_diffusion else LORA_WEIGHT_NAME_SAFE
-            else:
-                weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME if is_custom_diffusion else LORA_WEIGHT_NAME
-
-        # Save the model
-        save_function(state_dict, os.path.join(save_directory, weight_name))
-        logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
-
-    def fuse_lora(self, lora_scale=1.0, safe_fusing=False):
-        self.lora_scale = lora_scale
-        self._safe_fusing = safe_fusing
-        self.apply(self._fuse_lora_apply)
-
-    def _fuse_lora_apply(self, module):
-        if not USE_PEFT_BACKEND:
-            if hasattr(module, "_fuse_lora"):
-                module._fuse_lora(self.lora_scale, self._safe_fusing)
-        else:
-            from peft.tuners.tuners_utils import BaseTunerLayer
-
-            if isinstance(module, BaseTunerLayer):
-                if self.lora_scale != 1.0:
-                    module.scale_layer(self.lora_scale)
-                module.merge(safe_merge=self._safe_fusing)
-
-    def unfuse_lora(self):
-        self.apply(self._unfuse_lora_apply)
-
-    def _unfuse_lora_apply(self, module):
-        if not USE_PEFT_BACKEND:
-            if hasattr(module, "_unfuse_lora"):
-                module._unfuse_lora()
-        else:
-            from peft.tuners.tuners_utils import BaseTunerLayer
-
-            if isinstance(module, BaseTunerLayer):
-                module.unmerge()
-
-    def set_adapters(
-        self,
-        adapter_names: Union[List[str], str],
-        weights: Optional[Union[List[float], float]] = None,
-    ):
-        """
-        Sets the adapter layers for the unet.
-
-        Args:
-            adapter_names (`List[str]` or `str`):
-                The names of the adapters to use.
-            weights (`Union[List[float], float]`, *optional*):
-                The adapter(s) weights to use with the UNet. If `None`, the weights are set to `1.0` for all the
-                adapters.
-        """
-        if not USE_PEFT_BACKEND:
-            raise ValueError("PEFT backend is required for `set_adapters()`.")
-
-        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
-
-        if weights is None:
-            weights = [1.0] * len(adapter_names)
-        elif isinstance(weights, float):
-            weights = [weights] * len(adapter_names)
-
-        if len(adapter_names) != len(weights):
-            raise ValueError(
-                f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
-            )
-
-        set_weights_and_activate_adapters(self, adapter_names, weights)
-
-    def disable_lora(self):
-        """
-        Disables the active LoRA layers for the unet.
-        """
-        if not USE_PEFT_BACKEND:
-            raise ValueError("PEFT backend is required for this method.")
-        set_adapter_layers(self, enabled=False)
-
-    def enable_lora(self):
-        """
-        Enables the active LoRA layers for the unet.
-        """
-        if not USE_PEFT_BACKEND:
-            raise ValueError("PEFT backend is required for this method.")
-        set_adapter_layers(self, enabled=True)
-
-    def delete_adapters(self, adapter_names: Union[List[str], str]):
-        """
-        Deletes the LoRA layers of `adapter_name` for the unet.
-
-        Args:
-            adapter_names (`Union[List[str], str]`):
-                The names of the adapter to delete. Can be a single string or a list of strings
-        """
-        if not USE_PEFT_BACKEND:
-            raise ValueError("PEFT backend is required for this method.")
-
-        if isinstance(adapter_names, str):
-            adapter_names = [adapter_names]
-
-        for adapter_name in adapter_names:
-            delete_adapter_layers(self, adapter_name)
-
-            # Pop also the corresponding adapter from the config
-            if hasattr(self, "peft_config"):
-                self.peft_config.pop(adapter_name, None)
-
-
-def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
-    cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
-    force_download = kwargs.pop("force_download", False)
-    resume_download = kwargs.pop("resume_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)
-    subfolder = kwargs.pop("subfolder", None)
-    weight_name = kwargs.pop("weight_name", None)
-    use_safetensors = kwargs.pop("use_safetensors", None)
-
-    allow_pickle = False
-    if use_safetensors is None:
-        use_safetensors = True
-        allow_pickle = True
-
-    user_agent = {
-        "file_type": "text_inversion",
-        "framework": "pytorch",
-    }
-    state_dicts = []
-    for pretrained_model_name_or_path in pretrained_model_name_or_paths:
-        if not isinstance(pretrained_model_name_or_path, (dict, torch.Tensor)):
-            # 3.1. Load textual inversion file
-            model_file = None
-
-            # Let's first try to load .safetensors weights
-            if (use_safetensors and weight_name is None) or (
-                weight_name is not None and weight_name.endswith(".safetensors")
-            ):
-                try:
-                    model_file = _get_model_file(
-                        pretrained_model_name_or_path,
-                        weights_name=weight_name or TEXT_INVERSION_NAME_SAFE,
-                        cache_dir=cache_dir,
-                        force_download=force_download,
-                        resume_download=resume_download,
-                        proxies=proxies,
-                        local_files_only=local_files_only,
-                        use_auth_token=use_auth_token,
-                        revision=revision,
-                        subfolder=subfolder,
-                        user_agent=user_agent,
-                    )
-                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
-                except Exception as e:
-                    if not allow_pickle:
-                        raise e
-
-                    model_file = None
-
-            if model_file is None:
-                model_file = _get_model_file(
-                    pretrained_model_name_or_path,
-                    weights_name=weight_name or TEXT_INVERSION_NAME,
-                    cache_dir=cache_dir,
-                    force_download=force_download,
-                    resume_download=resume_download,
-                    proxies=proxies,
-                    local_files_only=local_files_only,
-                    use_auth_token=use_auth_token,
-                    revision=revision,
-                    subfolder=subfolder,
-                    user_agent=user_agent,
-                )
-                state_dict = torch.load(model_file, map_location="cpu")
-        else:
-            state_dict = pretrained_model_name_or_path
-
-        state_dicts.append(state_dict)
-
-    return state_dicts
-
-
-class TextualInversionLoaderMixin:
-    r"""
-    Load textual inversion tokens and embeddings to the tokenizer and text encoder.
-    """
-
-    def maybe_convert_prompt(self, prompt: Union[str, List[str]], tokenizer: "PreTrainedTokenizer"):  # noqa: F821
-        r"""
-        Processes prompts that include a special token corresponding to a multi-vector textual inversion embedding to
-        be replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
-        inversion token or if the textual inversion token is a single vector, the input prompt is returned.
-
-        Parameters:
-            prompt (`str` or list of `str`):
-                The prompt or prompts to guide the image generation.
-            tokenizer (`PreTrainedTokenizer`):
-                The tokenizer responsible for encoding the prompt into input tokens.
-
-        Returns:
-            `str` or list of `str`: The converted prompt
-        """
-        if not isinstance(prompt, List):
-            prompts = [prompt]
-        else:
-            prompts = prompt
-
-        prompts = [self._maybe_convert_prompt(p, tokenizer) for p in prompts]
-
-        if not isinstance(prompt, List):
-            return prompts[0]
-
-        return prompts
-
-    def _maybe_convert_prompt(self, prompt: str, tokenizer: "PreTrainedTokenizer"):  # noqa: F821
-        r"""
-        Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
-        to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
-        is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
-        inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
-
-        Parameters:
-            prompt (`str`):
-                The prompt to guide the image generation.
-            tokenizer (`PreTrainedTokenizer`):
-                The tokenizer responsible for encoding the prompt into input tokens.
-
-        Returns:
-            `str`: The converted prompt
-        """
-        tokens = tokenizer.tokenize(prompt)
-        unique_tokens = set(tokens)
-        for token in unique_tokens:
-            if token in tokenizer.added_tokens_encoder:
-                replacement = token
-                i = 1
-                while f"{token}_{i}" in tokenizer.added_tokens_encoder:
-                    replacement += f" {token}_{i}"
-                    i += 1
-
-                prompt = prompt.replace(token, replacement)
-
-        return prompt
-
-    def _check_text_inv_inputs(self, tokenizer, text_encoder, pretrained_model_name_or_paths, tokens):
-        if tokenizer is None:
-            raise ValueError(
-                f"{self.__class__.__name__} requires `self.tokenizer` or passing a `tokenizer` of type `PreTrainedTokenizer` for calling"
-                f" `{self.load_textual_inversion.__name__}`"
-            )
-
-        if text_encoder is None:
-            raise ValueError(
-                f"{self.__class__.__name__} requires `self.text_encoder` or passing a `text_encoder` of type `PreTrainedModel` for calling"
-                f" `{self.load_textual_inversion.__name__}`"
-            )
-
-        if len(pretrained_model_name_or_paths) != len(tokens):
-            raise ValueError(
-                f"You have passed a list of models of length {len(pretrained_model_name_or_paths)}, and list of tokens of length {len(tokens)} "
-                f"Make sure both lists have the same length."
-            )
-
-        valid_tokens = [t for t in tokens if t is not None]
-        if len(set(valid_tokens)) < len(valid_tokens):
-            raise ValueError(f"You have passed a list of tokens that contains duplicates: {tokens}")
-
-    @staticmethod
-    def _retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer):
-        all_tokens = []
-        all_embeddings = []
-        for state_dict, token in zip(state_dicts, tokens):
-            if isinstance(state_dict, torch.Tensor):
-                if token is None:
-                    raise ValueError(
-                        "You are trying to load a textual inversion embedding that has been saved as a PyTorch tensor. Make sure to pass the name of the corresponding token in this case: `token=...`."
-                    )
-                loaded_token = token
-                embedding = state_dict
-            elif len(state_dict) == 1:
-                # diffusers
-                loaded_token, embedding = next(iter(state_dict.items()))
-            elif "string_to_param" in state_dict:
-                # A1111
-                loaded_token = state_dict["name"]
-                embedding = state_dict["string_to_param"]["*"]
-            else:
-                raise ValueError(
-                    f"Loaded state dictonary is incorrect: {state_dict}. \n\n"
-                    "Please verify that the loaded state dictionary of the textual embedding either only has a single key or includes the `string_to_param`"
-                    " input key."
-                )
-
-            if token is not None and loaded_token != token:
-                logger.info(f"The loaded token: {loaded_token} is overwritten by the passed token {token}.")
-            else:
-                token = loaded_token
-
-            if token in tokenizer.get_vocab():
-                raise ValueError(
-                    f"Token {token} already in tokenizer vocabulary. Please choose a different token name or remove {token} and embedding from the tokenizer and text encoder."
-                )
-
-            all_tokens.append(token)
-            all_embeddings.append(embedding)
-
-        return all_tokens, all_embeddings
-
-    @staticmethod
-    def _extend_tokens_and_embeddings(tokens, embeddings, tokenizer):
-        all_tokens = []
-        all_embeddings = []
-
-        for embedding, token in zip(embeddings, tokens):
-            if f"{token}_1" in tokenizer.get_vocab():
-                multi_vector_tokens = [token]
-                i = 1
-                while f"{token}_{i}" in tokenizer.added_tokens_encoder:
-                    multi_vector_tokens.append(f"{token}_{i}")
-                    i += 1
-
-                raise ValueError(
-                    f"Multi-vector Token {multi_vector_tokens} already in tokenizer vocabulary. Please choose a different token name or remove the {multi_vector_tokens} and embedding from the tokenizer and text encoder."
-                )
-
-            is_multi_vector = len(embedding.shape) > 1 and embedding.shape[0] > 1
-            if is_multi_vector:
-                all_tokens += [token] + [f"{token}_{i}" for i in range(1, embedding.shape[0])]
-                all_embeddings += [e for e in embedding]  # noqa: C416
-            else:
-                all_tokens += [token]
-                all_embeddings += [embedding[0]] if len(embedding.shape) > 1 else [embedding]
-
-        return all_tokens, all_embeddings
-
-    def load_textual_inversion(
-        self,
-        pretrained_model_name_or_path: Union[str, List[str], Dict[str, torch.Tensor], List[Dict[str, torch.Tensor]]],
-        token: Optional[Union[str, List[str]]] = None,
-        tokenizer: Optional["PreTrainedTokenizer"] = None,  # noqa: F821
-        text_encoder: Optional["PreTrainedModel"] = None,  # noqa: F821
-        **kwargs,
-    ):
-        r"""
-        Load textual inversion embeddings into the text encoder of [`StableDiffusionPipeline`] (both 🤗 Diffusers and
-        Automatic1111 formats are supported).
-
-        Parameters:
-            pretrained_model_name_or_path (`str` or `os.PathLike` or `List[str or os.PathLike]` or `Dict` or `List[Dict]`):
-                Can be either one of the following or a list of them:
-
-                    - A string, the *model id* (for example `sd-concepts-library/low-poly-hd-logos-icons`) of a
-                      pretrained model hosted on the Hub.
-                    - A path to a *directory* (for example `./my_text_inversion_directory/`) containing the textual
-                      inversion weights.
-                    - A path to a *file* (for example `./my_text_inversions.pt`) containing textual inversion weights.
-                    - A [torch state
-                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
-
-            token (`str` or `List[str]`, *optional*):
-                Override the token to use for the textual inversion weights. If `pretrained_model_name_or_path` is a
-                list, then `token` must also be a list of equal length.
-            text_encoder ([`~transformers.CLIPTextModel`], *optional*):
-                Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
-                If not specified, function will take self.tokenizer.
-            tokenizer ([`~transformers.CLIPTokenizer`], *optional*):
-                A `CLIPTokenizer` to tokenize text. If not specified, function will take self.tokenizer.
-            weight_name (`str`, *optional*):
-                Name of a custom weight file. This should be used when:
-
-                    - The saved textual inversion file is in 🤗 Diffusers format, but was saved under a specific weight
-                      name such as `text_inv.bin`.
-                    - The saved textual inversion file is in the Automatic1111 format.
-            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.
-            force_download (`bool`, *optional*, defaults to `False`):
-                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
-                cached versions if they exist.
-            resume_download (`bool`, *optional*, defaults to `False`):
-                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.
-            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.
-            subfolder (`str`, *optional*, defaults to `""`):
-                The subfolder location of a model file within a larger model repository on the Hub or locally.
-            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.
-
-        Example:
-
-        To load a textual inversion embedding vector in 🤗 Diffusers format:
-
-        ```py
-        from diffusers import StableDiffusionPipeline
-        import torch
-
-        model_id = "runwayml/stable-diffusion-v1-5"
-        pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
-
-        pipe.load_textual_inversion("sd-concepts-library/cat-toy")
-
-        prompt = "A <cat-toy> backpack"
-
-        image = pipe(prompt, num_inference_steps=50).images[0]
-        image.save("cat-backpack.png")
-        ```
+    from ..models.lora import PatchedLoraProjection, text_encoder_attn_modules, text_encoder_mlp_modules
 
-        To load a textual inversion embedding vector in Automatic1111 format, make sure to download the vector first
-        (for example from [civitAI](https://civitai.com/models/3036?modelVersionId=9857)) and then load the vector
-        locally:
-
-        ```py
-        from diffusers import StableDiffusionPipeline
-        import torch
-
-        model_id = "runwayml/stable-diffusion-v1-5"
-        pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
-
-        pipe.load_textual_inversion("./charturnerv2.pt", token="charturnerv2")
-
-        prompt = "charturnerv2, multiple views of the same character in the same outfit, a character turnaround of a woman wearing a black jacket and red shirt, best quality, intricate details."
-
-        image = pipe(prompt, num_inference_steps=50).images[0]
-        image.save("character.png")
-        ```
-
-        """
-        # 1. Set correct tokenizer and text encoder
-        tokenizer = tokenizer or getattr(self, "tokenizer", None)
-        text_encoder = text_encoder or getattr(self, "text_encoder", None)
-
-        # 2. Normalize inputs
-        pretrained_model_name_or_paths = (
-            [pretrained_model_name_or_path]
-            if not isinstance(pretrained_model_name_or_path, list)
-            else pretrained_model_name_or_path
-        )
-        tokens = len(pretrained_model_name_or_paths) * [token] if (isinstance(token, str) or token is None) else token
-
-        # 3. Check inputs
-        self._check_text_inv_inputs(tokenizer, text_encoder, pretrained_model_name_or_paths, tokens)
-
-        # 4. Load state dicts of textual embeddings
-        state_dicts = load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
-
-        # 4. Retrieve tokens and embeddings
-        tokens, embeddings = self._retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer)
-
-        # 5. Extend tokens and embeddings for multi vector
-        tokens, embeddings = self._extend_tokens_and_embeddings(tokens, embeddings, tokenizer)
-
-        # 6. Make sure all embeddings have the correct size
-        expected_emb_dim = text_encoder.get_input_embeddings().weight.shape[-1]
-        if any(expected_emb_dim != emb.shape[-1] for emb in embeddings):
-            raise ValueError(
-                "Loaded embeddings are of incorrect shape. Expected each textual inversion embedding "
-                "to be of shape {input_embeddings.shape[-1]}, but are {embeddings.shape[-1]} "
-            )
-
-        # 7. Now we can be sure that loading the embedding matrix works
-        # < Unsafe code:
-
-        # 7.1 Offload all hooks in case the pipeline was cpu offloaded before make sure, we offload and onload again
-        is_model_cpu_offload = False
-        is_sequential_cpu_offload = False
-        for _, component in self.components.items():
-            if isinstance(component, nn.Module):
-                if hasattr(component, "_hf_hook"):
-                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
-                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
-                    logger.info(
-                        "Accelerate hooks detected. Since you have called `load_textual_inversion()`, the previous hooks will be first removed. Then the textual inversion parameters will be loaded and the hooks will be applied again."
-                    )
-                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
-
-        # 7.2 save expected device and dtype
-        device = text_encoder.device
-        dtype = text_encoder.dtype
-
-        # 7.3 Increase token embedding matrix
-        text_encoder.resize_token_embeddings(len(tokenizer) + len(tokens))
-        input_embeddings = text_encoder.get_input_embeddings().weight
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
 
-        # 7.4 Load token and embedding
-        for token, embedding in zip(tokens, embeddings):
-            # add tokens and get ids
-            tokenizer.add_tokens(token)
-            token_id = tokenizer.convert_tokens_to_ids(token)
-            input_embeddings.data[token_id] = embedding
-            logger.info(f"Loaded textual inversion embedding for {token}.")
+logger = logging.get_logger(__name__)
 
-        input_embeddings.to(dtype=dtype, device=device)
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
 
-        # 7.5 Offload the model again
-        if is_model_cpu_offload:
-            self.enable_model_cpu_offload()
-        elif is_sequential_cpu_offload:
-            self.enable_sequential_cpu_offload()
+LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
 
-        # / Unsafe Code >
+LORA_DEPRECATION_MESSAGE = "You are using an old version of LoRA backend. This will be deprecated in the next releases in favor of PEFT make sure to install the latest PEFT and transformers packages in the future."
 
 
 class LoraLoaderMixin:
@@ -2531,9 +1414,8 @@ class LoraLoaderMixin:
 
     def delete_adapters(self, adapter_names: Union[List[str], str]):
         """
-        Deletes the LoRA layers of `adapter_name` for the unet and text-encoder(s).
-
         Args:
+        Deletes the LoRA layers of `adapter_name` for the unet and text-encoder(s).
             adapter_names (`Union[List[str], str]`):
                 The names of the adapter to delete. Can be a single string or a list of strings
         """
@@ -2647,588 +1529,6 @@ class LoraLoaderMixin:
                         text_encoder_module.lora_B[adapter_name].to(device)
 
 
-class FromSingleFileMixin:
-    """
-    Load model weights saved in the `.ckpt` format into a [`DiffusionPipeline`].
-    """
-
-    @classmethod
-    def from_ckpt(cls, *args, **kwargs):
-        deprecation_message = "The function `from_ckpt` is deprecated in favor of `from_single_file` and will be removed in diffusers v.0.21. Please make sure to use `StableDiffusionPipeline.from_single_file(...)` instead."
-        deprecate("from_ckpt", "0.21.0", deprecation_message, standard_warn=False)
-        return cls.from_single_file(*args, **kwargs)
-
-    @classmethod
-    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
-        r"""
-        Instantiate a [`DiffusionPipeline`] from pretrained pipeline weights saved in the `.ckpt` or `.safetensors`
-        format. The pipeline is set in evaluation mode (`model.eval()`) by default.
-
-        Parameters:
-            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
-                Can be either:
-                    - A link to the `.ckpt` file (for example
-                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
-                    - A path to a *file* containing all pipeline weights.
-            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.
-            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.
-            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.
-            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.
-            extract_ema (`bool`, *optional*, defaults to `False`):
-                Whether to extract the EMA weights or not. Pass `True` to extract the EMA weights which usually yield
-                higher quality images for inference. Non-EMA weights are usually better for continuing finetuning.
-            upcast_attention (`bool`, *optional*, defaults to `None`):
-                Whether the attention computation should always be upcasted.
-            image_size (`int`, *optional*, defaults to 512):
-                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
-                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
-            prediction_type (`str`, *optional*):
-                The prediction type the model was trained on. Use `'epsilon'` for all Stable Diffusion v1 models and
-                the Stable Diffusion v2 base model. Use `'v_prediction'` for Stable Diffusion v2.
-            num_in_channels (`int`, *optional*, defaults to `None`):
-                The number of input channels. If `None`, it is automatically inferred.
-            scheduler_type (`str`, *optional*, defaults to `"pndm"`):
-                Type of scheduler to use. Should be one of `["pndm", "lms", "heun", "euler", "euler-ancestral", "dpm",
-                "ddim"]`.
-            load_safety_checker (`bool`, *optional*, defaults to `True`):
-                Whether to load the safety checker or not.
-            text_encoder ([`~transformers.CLIPTextModel`], *optional*, defaults to `None`):
-                An instance of `CLIPTextModel` to use, specifically the
-                [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. If this
-                parameter is `None`, the function loads a new instance of `CLIPTextModel` by itself if needed.
-            vae (`AutoencoderKL`, *optional*, defaults to `None`):
-                Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. If
-                this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed.
-            tokenizer ([`~transformers.CLIPTokenizer`], *optional*, defaults to `None`):
-                An instance of `CLIPTokenizer` to use. If this parameter is `None`, the function loads a new instance
-                of `CLIPTokenizer` by itself if needed.
-            original_config_file (`str`):
-                Path to `.yaml` config file corresponding to the original architecture. If `None`, will be
-                automatically inferred by looking for a key that only exists in SD2.0 models.
-            kwargs (remaining dictionary of keyword arguments, *optional*):
-                Can be used to overwrite load and saveable variables (for example the pipeline components of the
-                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
-                method. See example below for more information.
-
-        Examples:
-
-        ```py
-        >>> from diffusers import StableDiffusionPipeline
-
-        >>> # Download pipeline from huggingface.co and cache.
-        >>> pipeline = StableDiffusionPipeline.from_single_file(
-        ...     "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix.safetensors"
-        ... )
-
-        >>> # Download pipeline from local file
-        >>> # file is downloaded under ./v1-5-pruned-emaonly.ckpt
-        >>> pipeline = StableDiffusionPipeline.from_single_file("./v1-5-pruned-emaonly")
-
-        >>> # Enable float16 and move to GPU
-        >>> pipeline = StableDiffusionPipeline.from_single_file(
-        ...     "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt",
-        ...     torch_dtype=torch.float16,
-        ... )
-        >>> pipeline.to("cuda")
-        ```
-        """
-        # import here to avoid circular dependency
-        from .pipelines.stable_diffusion.convert_from_ckpt import download_from_original_stable_diffusion_ckpt
-
-        original_config_file = kwargs.pop("original_config_file", None)
-        config_files = kwargs.pop("config_files", None)
-        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)
-        extract_ema = kwargs.pop("extract_ema", False)
-        image_size = kwargs.pop("image_size", None)
-        scheduler_type = kwargs.pop("scheduler_type", "pndm")
-        num_in_channels = kwargs.pop("num_in_channels", None)
-        upcast_attention = kwargs.pop("upcast_attention", None)
-        load_safety_checker = kwargs.pop("load_safety_checker", True)
-        prediction_type = kwargs.pop("prediction_type", None)
-        text_encoder = kwargs.pop("text_encoder", None)
-        vae = kwargs.pop("vae", None)
-        controlnet = kwargs.pop("controlnet", None)
-        adapter = kwargs.pop("adapter", None)
-        tokenizer = kwargs.pop("tokenizer", None)
-
-        torch_dtype = kwargs.pop("torch_dtype", None)
-
-        use_safetensors = kwargs.pop("use_safetensors", None)
-
-        pipeline_name = cls.__name__
-        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
-        from_safetensors = file_extension == "safetensors"
-
-        if from_safetensors and use_safetensors is False:
-            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
-
-        # TODO: For now we only support stable diffusion
-        stable_unclip = None
-        model_type = None
-
-        if pipeline_name in [
-            "StableDiffusionControlNetPipeline",
-            "StableDiffusionControlNetImg2ImgPipeline",
-            "StableDiffusionControlNetInpaintPipeline",
-        ]:
-            from .models.controlnet import ControlNetModel
-            from .pipelines.controlnet.multicontrolnet import MultiControlNetModel
-
-            #  list/tuple or a single instance of ControlNetModel or MultiControlNetModel
-            if not (
-                isinstance(controlnet, (ControlNetModel, MultiControlNetModel))
-                or isinstance(controlnet, (list, tuple))
-                and isinstance(controlnet[0], ControlNetModel)
-            ):
-                raise ValueError("ControlNet needs to be passed if loading from ControlNet pipeline.")
-        elif "StableDiffusion" in pipeline_name:
-            # Model type will be inferred from the checkpoint.
-            pass
-        elif pipeline_name == "StableUnCLIPPipeline":
-            model_type = "FrozenOpenCLIPEmbedder"
-            stable_unclip = "txt2img"
-        elif pipeline_name == "StableUnCLIPImg2ImgPipeline":
-            model_type = "FrozenOpenCLIPEmbedder"
-            stable_unclip = "img2img"
-        elif pipeline_name == "PaintByExamplePipeline":
-            model_type = "PaintByExample"
-        elif pipeline_name == "LDMTextToImagePipeline":
-            model_type = "LDMTextToImage"
-        else:
-            raise ValueError(f"Unhandled pipeline class: {pipeline_name}")
-
-        # remove huggingface url
-        has_valid_url_prefix = False
-        valid_url_prefixes = ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]
-        for prefix in valid_url_prefixes:
-            if pretrained_model_link_or_path.startswith(prefix):
-                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
-                has_valid_url_prefix = True
-
-        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
-        ckpt_path = Path(pretrained_model_link_or_path)
-        if not ckpt_path.is_file():
-            if not has_valid_url_prefix:
-                raise ValueError(
-                    f"The provided path is either not a file or a valid huggingface URL was not provided. Valid URLs begin with {', '.join(valid_url_prefixes)}"
-                )
-
-            # get repo_id and (potentially nested) file path of ckpt in repo
-            repo_id = "/".join(ckpt_path.parts[:2])
-            file_path = "/".join(ckpt_path.parts[2:])
-
-            if file_path.startswith("blob/"):
-                file_path = file_path[len("blob/") :]
-
-            if file_path.startswith("main/"):
-                file_path = file_path[len("main/") :]
-
-            pretrained_model_link_or_path = hf_hub_download(
-                repo_id,
-                filename=file_path,
-                cache_dir=cache_dir,
-                resume_download=resume_download,
-                proxies=proxies,
-                local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
-                revision=revision,
-                force_download=force_download,
-            )
-
-        pipe = download_from_original_stable_diffusion_ckpt(
-            pretrained_model_link_or_path,
-            pipeline_class=cls,
-            model_type=model_type,
-            stable_unclip=stable_unclip,
-            controlnet=controlnet,
-            adapter=adapter,
-            from_safetensors=from_safetensors,
-            extract_ema=extract_ema,
-            image_size=image_size,
-            scheduler_type=scheduler_type,
-            num_in_channels=num_in_channels,
-            upcast_attention=upcast_attention,
-            load_safety_checker=load_safety_checker,
-            prediction_type=prediction_type,
-            text_encoder=text_encoder,
-            vae=vae,
-            tokenizer=tokenizer,
-            original_config_file=original_config_file,
-            config_files=config_files,
-            local_files_only=local_files_only,
-        )
-
-        if torch_dtype is not None:
-            pipe.to(torch_dtype=torch_dtype)
-
-        return pipe
-
-
-class FromOriginalVAEMixin:
-    @classmethod
-    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
-        r"""
-        Instantiate a [`AutoencoderKL`] from pretrained controlnet weights saved in the original `.ckpt` or
-        `.safetensors` format. The pipeline is format. The pipeline is set in evaluation mode (`model.eval()`) by
-        default.
-
-        Parameters:
-            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
-                Can be either:
-                    - A link to the `.ckpt` file (for example
-                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
-                    - A path to a *file* containing all pipeline weights.
-            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.
-            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.
-            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.
-            image_size (`int`, *optional*, defaults to 512):
-                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
-                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
-            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.
-            upcast_attention (`bool`, *optional*, defaults to `None`):
-                Whether the attention computation should always be upcasted.
-            scaling_factor (`float`, *optional*, defaults to 0.18215):
-                The component-wise standard deviation of the trained latent space computed using the first batch of the
-                training set. This is used to scale the latent space to have unit variance when training the diffusion
-                model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
-                diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z
-                = 1 / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution
-                Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
-            kwargs (remaining dictionary of keyword arguments, *optional*):
-                Can be used to overwrite load and saveable variables (for example the pipeline components of the
-                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
-                method. See example below for more information.
-
-        <Tip warning={true}>
-
-            Make sure to pass both `image_size` and `scaling_factor` to `from_single_file()` if you want to load
-            a VAE that does accompany a stable diffusion model of v2 or higher or SDXL.
-
-        </Tip>
-
-        Examples:
-
-        ```py
-        from diffusers import AutoencoderKL
-
-        url = "https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors"  # can also be local file
-        model = AutoencoderKL.from_single_file(url)
-        ```
-        """
-        if not is_omegaconf_available():
-            raise ValueError(BACKENDS_MAPPING["omegaconf"][1])
-
-        from omegaconf import OmegaConf
-
-        from .models import AutoencoderKL
-
-        # import here to avoid circular dependency
-        from .pipelines.stable_diffusion.convert_from_ckpt import (
-            convert_ldm_vae_checkpoint,
-            create_vae_diffusers_config,
-        )
-
-        config_file = kwargs.pop("config_file", None)
-        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)
-        image_size = kwargs.pop("image_size", None)
-        scaling_factor = kwargs.pop("scaling_factor", None)
-        kwargs.pop("upcast_attention", None)
-
-        torch_dtype = kwargs.pop("torch_dtype", None)
-
-        use_safetensors = kwargs.pop("use_safetensors", None)
-
-        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
-        from_safetensors = file_extension == "safetensors"
-
-        if from_safetensors and use_safetensors is False:
-            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
-
-        # remove huggingface url
-        for prefix in ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]:
-            if pretrained_model_link_or_path.startswith(prefix):
-                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
-
-        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
-        ckpt_path = Path(pretrained_model_link_or_path)
-        if not ckpt_path.is_file():
-            # get repo_id and (potentially nested) file path of ckpt in repo
-            repo_id = "/".join(ckpt_path.parts[:2])
-            file_path = "/".join(ckpt_path.parts[2:])
-
-            if file_path.startswith("blob/"):
-                file_path = file_path[len("blob/") :]
-
-            if file_path.startswith("main/"):
-                file_path = file_path[len("main/") :]
-
-            pretrained_model_link_or_path = hf_hub_download(
-                repo_id,
-                filename=file_path,
-                cache_dir=cache_dir,
-                resume_download=resume_download,
-                proxies=proxies,
-                local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
-                revision=revision,
-                force_download=force_download,
-            )
-
-        if from_safetensors:
-            from safetensors import safe_open
-
-            checkpoint = {}
-            with safe_open(pretrained_model_link_or_path, framework="pt", device="cpu") as f:
-                for key in f.keys():
-                    checkpoint[key] = f.get_tensor(key)
-        else:
-            checkpoint = torch.load(pretrained_model_link_or_path, map_location="cpu")
-
-        if "state_dict" in checkpoint:
-            checkpoint = checkpoint["state_dict"]
-
-        if config_file is None:
-            config_url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml"
-            config_file = BytesIO(requests.get(config_url).content)
-
-        original_config = OmegaConf.load(config_file)
-
-        # default to sd-v1-5
-        image_size = image_size or 512
-
-        vae_config = create_vae_diffusers_config(original_config, image_size=image_size)
-        converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
-
-        if scaling_factor is None:
-            if (
-                "model" in original_config
-                and "params" in original_config.model
-                and "scale_factor" in original_config.model.params
-            ):
-                vae_scaling_factor = original_config.model.params.scale_factor
-            else:
-                vae_scaling_factor = 0.18215  # default SD scaling factor
-
-        vae_config["scaling_factor"] = vae_scaling_factor
-
-        ctx = init_empty_weights if is_accelerate_available() else nullcontext
-        with ctx():
-            vae = AutoencoderKL(**vae_config)
-
-        if is_accelerate_available():
-            load_model_dict_into_meta(vae, converted_vae_checkpoint, device="cpu")
-        else:
-            vae.load_state_dict(converted_vae_checkpoint)
-
-        if torch_dtype is not None:
-            vae.to(dtype=torch_dtype)
-
-        return vae
-
-
-class FromOriginalControlnetMixin:
-    @classmethod
-    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
-        r"""
-        Instantiate a [`ControlNetModel`] from pretrained controlnet weights saved in the original `.ckpt` or
-        `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
-
-        Parameters:
-            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
-                Can be either:
-                    - A link to the `.ckpt` file (for example
-                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
-                    - A path to a *file* containing all pipeline weights.
-            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.
-            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.
-            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.
-            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.
-            image_size (`int`, *optional*, defaults to 512):
-                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
-                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
-            upcast_attention (`bool`, *optional*, defaults to `None`):
-                Whether the attention computation should always be upcasted.
-            kwargs (remaining dictionary of keyword arguments, *optional*):
-                Can be used to overwrite load and saveable variables (for example the pipeline components of the
-                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
-                method. See example below for more information.
-
-        Examples:
-
-        ```py
-        from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
-
-        url = "https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_canny.pth"  # can also be a local path
-        model = ControlNetModel.from_single_file(url)
-
-        url = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.safetensors"  # can also be a local path
-        pipe = StableDiffusionControlNetPipeline.from_single_file(url, controlnet=controlnet)
-        ```
-        """
-        # import here to avoid circular dependency
-        from .pipelines.stable_diffusion.convert_from_ckpt import download_controlnet_from_original_ckpt
-
-        config_file = kwargs.pop("config_file", None)
-        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)
-        num_in_channels = kwargs.pop("num_in_channels", None)
-        use_linear_projection = kwargs.pop("use_linear_projection", None)
-        revision = kwargs.pop("revision", None)
-        extract_ema = kwargs.pop("extract_ema", False)
-        image_size = kwargs.pop("image_size", None)
-        upcast_attention = kwargs.pop("upcast_attention", None)
-
-        torch_dtype = kwargs.pop("torch_dtype", None)
-
-        use_safetensors = kwargs.pop("use_safetensors", None)
-
-        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
-        from_safetensors = file_extension == "safetensors"
-
-        if from_safetensors and use_safetensors is False:
-            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
-
-        # remove huggingface url
-        for prefix in ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]:
-            if pretrained_model_link_or_path.startswith(prefix):
-                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
-
-        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
-        ckpt_path = Path(pretrained_model_link_or_path)
-        if not ckpt_path.is_file():
-            # get repo_id and (potentially nested) file path of ckpt in repo
-            repo_id = "/".join(ckpt_path.parts[:2])
-            file_path = "/".join(ckpt_path.parts[2:])
-
-            if file_path.startswith("blob/"):
-                file_path = file_path[len("blob/") :]
-
-            if file_path.startswith("main/"):
-                file_path = file_path[len("main/") :]
-
-            pretrained_model_link_or_path = hf_hub_download(
-                repo_id,
-                filename=file_path,
-                cache_dir=cache_dir,
-                resume_download=resume_download,
-                proxies=proxies,
-                local_files_only=local_files_only,
-                use_auth_token=use_auth_token,
-                revision=revision,
-                force_download=force_download,
-            )
-
-        if config_file is None:
-            config_url = "https://raw.githubusercontent.com/lllyasviel/ControlNet/main/models/cldm_v15.yaml"
-            config_file = BytesIO(requests.get(config_url).content)
-
-        image_size = image_size or 512
-
-        controlnet = download_controlnet_from_original_ckpt(
-            pretrained_model_link_or_path,
-            original_config_file=config_file,
-            image_size=image_size,
-            extract_ema=extract_ema,
-            num_in_channels=num_in_channels,
-            upcast_attention=upcast_attention,
-            from_safetensors=from_safetensors,
-            use_linear_projection=use_linear_projection,
-        )
-
-        if torch_dtype is not None:
-            controlnet.to(dtype=torch_dtype)
-
-        return controlnet
-
-
 class StableDiffusionXLLoraLoaderMixin(LoraLoaderMixin):
     """This class overrides `LoraLoaderMixin` with LoRA loading/saving code that's specific to SDXL"""
 

src/diffusers/loaders/single_file.py

+# Copyright 2023 The HuggingFace Team. 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.
+from contextlib import nullcontext
+from io import BytesIO
+from pathlib import Path
+
+import requests
+import torch
+from huggingface_hub import hf_hub_download
+
+from ..utils import (
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
+    deprecate,
+    is_accelerate_available,
+    is_omegaconf_available,
+    is_transformers_available,
+    logging,
+)
+from ..utils.import_utils import BACKENDS_MAPPING
+
+
+if is_transformers_available():
+    pass
+
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+
+logger = logging.get_logger(__name__)
+
+
+class FromSingleFileMixin:
+    """
+    Load model weights saved in the `.ckpt` format into a [`DiffusionPipeline`].
+    """
+
+    @classmethod
+    def from_ckpt(cls, *args, **kwargs):
+        deprecation_message = "The function `from_ckpt` is deprecated in favor of `from_single_file` and will be removed in diffusers v.0.21. Please make sure to use `StableDiffusionPipeline.from_single_file(...)` instead."
+        deprecate("from_ckpt", "0.21.0", deprecation_message, standard_warn=False)
+        return cls.from_single_file(*args, **kwargs)
+
+    @classmethod
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`DiffusionPipeline`] from pretrained pipeline weights saved in the `.ckpt` or `.safetensors`
+        format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            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.
+            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.
+            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.
+            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.
+            extract_ema (`bool`, *optional*, defaults to `False`):
+                Whether to extract the EMA weights or not. Pass `True` to extract the EMA weights which usually yield
+                higher quality images for inference. Non-EMA weights are usually better for continuing finetuning.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            prediction_type (`str`, *optional*):
+                The prediction type the model was trained on. Use `'epsilon'` for all Stable Diffusion v1 models and
+                the Stable Diffusion v2 base model. Use `'v_prediction'` for Stable Diffusion v2.
+            num_in_channels (`int`, *optional*, defaults to `None`):
+                The number of input channels. If `None`, it is automatically inferred.
+            scheduler_type (`str`, *optional*, defaults to `"pndm"`):
+                Type of scheduler to use. Should be one of `["pndm", "lms", "heun", "euler", "euler-ancestral", "dpm",
+                "ddim"]`.
+            load_safety_checker (`bool`, *optional*, defaults to `True`):
+                Whether to load the safety checker or not.
+            text_encoder ([`~transformers.CLIPTextModel`], *optional*, defaults to `None`):
+                An instance of `CLIPTextModel` to use, specifically the
+                [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. If this
+                parameter is `None`, the function loads a new instance of `CLIPTextModel` by itself if needed.
+            vae (`AutoencoderKL`, *optional*, defaults to `None`):
+                Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. If
+                this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed.
+            tokenizer ([`~transformers.CLIPTokenizer`], *optional*, defaults to `None`):
+                An instance of `CLIPTokenizer` to use. If this parameter is `None`, the function loads a new instance
+                of `CLIPTokenizer` by itself if needed.
+            original_config_file (`str`):
+                Path to `.yaml` config file corresponding to the original architecture. If `None`, will be
+                automatically inferred by looking for a key that only exists in SD2.0 models.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        Examples:
+
+        ```py
+        >>> from diffusers import StableDiffusionPipeline
+
+        >>> # Download pipeline from huggingface.co and cache.
+        >>> pipeline = StableDiffusionPipeline.from_single_file(
+        ...     "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix.safetensors"
+        ... )
+
+        >>> # Download pipeline from local file
+        >>> # file is downloaded under ./v1-5-pruned-emaonly.ckpt
+        >>> pipeline = StableDiffusionPipeline.from_single_file("./v1-5-pruned-emaonly")
+
+        >>> # Enable float16 and move to GPU
+        >>> pipeline = StableDiffusionPipeline.from_single_file(
+        ...     "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt",
+        ...     torch_dtype=torch.float16,
+        ... )
+        >>> pipeline.to("cuda")
+        ```
+        """
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import download_from_original_stable_diffusion_ckpt
+
+        original_config_file = kwargs.pop("original_config_file", None)
+        config_files = kwargs.pop("config_files", None)
+        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)
+        extract_ema = kwargs.pop("extract_ema", False)
+        image_size = kwargs.pop("image_size", None)
+        scheduler_type = kwargs.pop("scheduler_type", "pndm")
+        num_in_channels = kwargs.pop("num_in_channels", None)
+        upcast_attention = kwargs.pop("upcast_attention", None)
+        load_safety_checker = kwargs.pop("load_safety_checker", True)
+        prediction_type = kwargs.pop("prediction_type", None)
+        text_encoder = kwargs.pop("text_encoder", None)
+        vae = kwargs.pop("vae", None)
+        controlnet = kwargs.pop("controlnet", None)
+        adapter = kwargs.pop("adapter", None)
+        tokenizer = kwargs.pop("tokenizer", None)
+
+        torch_dtype = kwargs.pop("torch_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        pipeline_name = cls.__name__
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        # TODO: For now we only support stable diffusion
+        stable_unclip = None
+        model_type = None
+
+        if pipeline_name in [
+            "StableDiffusionControlNetPipeline",
+            "StableDiffusionControlNetImg2ImgPipeline",
+            "StableDiffusionControlNetInpaintPipeline",
+        ]:
+            from .models.controlnet import ControlNetModel
+            from .pipelines.controlnet.multicontrolnet import MultiControlNetModel
+
+            #  list/tuple or a single instance of ControlNetModel or MultiControlNetModel
+            if not (
+                isinstance(controlnet, (ControlNetModel, MultiControlNetModel))
+                or isinstance(controlnet, (list, tuple))
+                and isinstance(controlnet[0], ControlNetModel)
+            ):
+                raise ValueError("ControlNet needs to be passed if loading from ControlNet pipeline.")
+        elif "StableDiffusion" in pipeline_name:
+            # Model type will be inferred from the checkpoint.
+            pass
+        elif pipeline_name == "StableUnCLIPPipeline":
+            model_type = "FrozenOpenCLIPEmbedder"
+            stable_unclip = "txt2img"
+        elif pipeline_name == "StableUnCLIPImg2ImgPipeline":
+            model_type = "FrozenOpenCLIPEmbedder"
+            stable_unclip = "img2img"
+        elif pipeline_name == "PaintByExamplePipeline":
+            model_type = "PaintByExample"
+        elif pipeline_name == "LDMTextToImagePipeline":
+            model_type = "LDMTextToImage"
+        else:
+            raise ValueError(f"Unhandled pipeline class: {pipeline_name}")
+
+        # remove huggingface url
+        has_valid_url_prefix = False
+        valid_url_prefixes = ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]
+        for prefix in valid_url_prefixes:
+            if pretrained_model_link_or_path.startswith(prefix):
+                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+                has_valid_url_prefix = True
+
+        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+        ckpt_path = Path(pretrained_model_link_or_path)
+        if not ckpt_path.is_file():
+            if not has_valid_url_prefix:
+                raise ValueError(
+                    f"The provided path is either not a file or a valid huggingface URL was not provided. Valid URLs begin with {', '.join(valid_url_prefixes)}"
+                )
+
+            # get repo_id and (potentially nested) file path of ckpt in repo
+            repo_id = "/".join(ckpt_path.parts[:2])
+            file_path = "/".join(ckpt_path.parts[2:])
+
+            if file_path.startswith("blob/"):
+                file_path = file_path[len("blob/") :]
+
+            if file_path.startswith("main/"):
+                file_path = file_path[len("main/") :]
+
+            pretrained_model_link_or_path = hf_hub_download(
+                repo_id,
+                filename=file_path,
+                cache_dir=cache_dir,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                force_download=force_download,
+            )
+
+        pipe = download_from_original_stable_diffusion_ckpt(
+            pretrained_model_link_or_path,
+            pipeline_class=cls,
+            model_type=model_type,
+            stable_unclip=stable_unclip,
+            controlnet=controlnet,
+            adapter=adapter,
+            from_safetensors=from_safetensors,
+            extract_ema=extract_ema,
+            image_size=image_size,
+            scheduler_type=scheduler_type,
+            num_in_channels=num_in_channels,
+            upcast_attention=upcast_attention,
+            load_safety_checker=load_safety_checker,
+            prediction_type=prediction_type,
+            text_encoder=text_encoder,
+            vae=vae,
+            tokenizer=tokenizer,
+            original_config_file=original_config_file,
+            config_files=config_files,
+            local_files_only=local_files_only,
+        )
+
+        if torch_dtype is not None:
+            pipe.to(torch_dtype=torch_dtype)
+
+        return pipe
+
+
+class FromOriginalVAEMixin:
+    @classmethod
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`AutoencoderKL`] from pretrained controlnet weights saved in the original `.ckpt` or
+        `.safetensors` format. The pipeline is format. The pipeline is set in evaluation mode (`model.eval()`) by
+        default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            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.
+            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.
+            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.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            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.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            scaling_factor (`float`, *optional*, defaults to 0.18215):
+                The component-wise standard deviation of the trained latent space computed using the first batch of the
+                training set. This is used to scale the latent space to have unit variance when training the diffusion
+                model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+                diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z
+                = 1 / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution
+                Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        <Tip warning={true}>
+
+            Make sure to pass both `image_size` and `scaling_factor` to `from_single_file()` if you want to load
+            a VAE that does accompany a stable diffusion model of v2 or higher or SDXL.
+
+        </Tip>
+
+        Examples:
+
+        ```py
+        from diffusers import AutoencoderKL
+
+        url = "https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors"  # can also be local file
+        model = AutoencoderKL.from_single_file(url)
+        ```
+        """
+        if not is_omegaconf_available():
+            raise ValueError(BACKENDS_MAPPING["omegaconf"][1])
+
+        from omegaconf import OmegaConf
+
+        from ..models import AutoencoderKL
+
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import (
+            convert_ldm_vae_checkpoint,
+            create_vae_diffusers_config,
+        )
+
+        config_file = kwargs.pop("config_file", None)
+        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)
+        image_size = kwargs.pop("image_size", None)
+        scaling_factor = kwargs.pop("scaling_factor", None)
+        kwargs.pop("upcast_attention", None)
+
+        torch_dtype = kwargs.pop("torch_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        # remove huggingface url
+        for prefix in ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]:
+            if pretrained_model_link_or_path.startswith(prefix):
+                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+
+        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+        ckpt_path = Path(pretrained_model_link_or_path)
+        if not ckpt_path.is_file():
+            # get repo_id and (potentially nested) file path of ckpt in repo
+            repo_id = "/".join(ckpt_path.parts[:2])
+            file_path = "/".join(ckpt_path.parts[2:])
+
+            if file_path.startswith("blob/"):
+                file_path = file_path[len("blob/") :]
+
+            if file_path.startswith("main/"):
+                file_path = file_path[len("main/") :]
+
+            pretrained_model_link_or_path = hf_hub_download(
+                repo_id,
+                filename=file_path,
+                cache_dir=cache_dir,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                force_download=force_download,
+            )
+
+        if from_safetensors:
+            from safetensors import safe_open
+
+            checkpoint = {}
+            with safe_open(pretrained_model_link_or_path, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    checkpoint[key] = f.get_tensor(key)
+        else:
+            checkpoint = torch.load(pretrained_model_link_or_path, map_location="cpu")
+
+        if "state_dict" in checkpoint:
+            checkpoint = checkpoint["state_dict"]
+
+        if config_file is None:
+            config_url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml"
+            config_file = BytesIO(requests.get(config_url).content)
+
+        original_config = OmegaConf.load(config_file)
+
+        # default to sd-v1-5
+        image_size = image_size or 512
+
+        vae_config = create_vae_diffusers_config(original_config, image_size=image_size)
+        converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
+
+        if scaling_factor is None:
+            if (
+                "model" in original_config
+                and "params" in original_config.model
+                and "scale_factor" in original_config.model.params
+            ):
+                vae_scaling_factor = original_config.model.params.scale_factor
+            else:
+                vae_scaling_factor = 0.18215  # default SD scaling factor
+
+        vae_config["scaling_factor"] = vae_scaling_factor
+
+        ctx = init_empty_weights if is_accelerate_available() else nullcontext
+        with ctx():
+            vae = AutoencoderKL(**vae_config)
+
+        if is_accelerate_available():
+            from ..models.modeling_utils import load_model_dict_into_meta
+
+            load_model_dict_into_meta(vae, converted_vae_checkpoint, device="cpu")
+        else:
+            vae.load_state_dict(converted_vae_checkpoint)
+
+        if torch_dtype is not None:
+            vae.to(dtype=torch_dtype)
+
+        return vae
+
+
+class FromOriginalControlnetMixin:
+    @classmethod
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`ControlNetModel`] from pretrained controlnet weights saved in the original `.ckpt` or
+        `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            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.
+            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.
+            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.
+            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.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        Examples:
+
+        ```py
+        from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
+
+        url = "https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_canny.pth"  # can also be a local path
+        model = ControlNetModel.from_single_file(url)
+
+        url = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.safetensors"  # can also be a local path
+        pipe = StableDiffusionControlNetPipeline.from_single_file(url, controlnet=controlnet)
+        ```
+        """
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import download_controlnet_from_original_ckpt
+
+        config_file = kwargs.pop("config_file", None)
+        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)
+        num_in_channels = kwargs.pop("num_in_channels", None)
+        use_linear_projection = kwargs.pop("use_linear_projection", None)
+        revision = kwargs.pop("revision", None)
+        extract_ema = kwargs.pop("extract_ema", False)
+        image_size = kwargs.pop("image_size", None)
+        upcast_attention = kwargs.pop("upcast_attention", None)
+
+        torch_dtype = kwargs.pop("torch_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        # remove huggingface url
+        for prefix in ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]:
+            if pretrained_model_link_or_path.startswith(prefix):
+                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+
+        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+        ckpt_path = Path(pretrained_model_link_or_path)
+        if not ckpt_path.is_file():
+            # get repo_id and (potentially nested) file path of ckpt in repo
+            repo_id = "/".join(ckpt_path.parts[:2])
+            file_path = "/".join(ckpt_path.parts[2:])
+
+            if file_path.startswith("blob/"):
+                file_path = file_path[len("blob/") :]
+
+            if file_path.startswith("main/"):
+                file_path = file_path[len("main/") :]
+
+            pretrained_model_link_or_path = hf_hub_download(
+                repo_id,
+                filename=file_path,
+                cache_dir=cache_dir,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                force_download=force_download,
+            )
+
+        if config_file is None:
+            config_url = "https://raw.githubusercontent.com/lllyasviel/ControlNet/main/models/cldm_v15.yaml"
+            config_file = BytesIO(requests.get(config_url).content)
+
+        image_size = image_size or 512
+
+        controlnet = download_controlnet_from_original_ckpt(
+            pretrained_model_link_or_path,
+            original_config_file=config_file,
+            image_size=image_size,
+            extract_ema=extract_ema,
+            num_in_channels=num_in_channels,
+            upcast_attention=upcast_attention,
+            from_safetensors=from_safetensors,
+            use_linear_projection=use_linear_projection,
+        )
+
+        if torch_dtype is not None:
+            controlnet.to(dtype=torch_dtype)
+
+        return controlnet

src/diffusers/loaders/textual_inversion.py

+# Copyright 2023 The HuggingFace Team. 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.
+from typing import Dict, List, Optional, Union
+
+import safetensors
+import torch
+from torch import nn
+
+from ..utils import (
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
+    _get_model_file,
+    is_accelerate_available,
+    is_transformers_available,
+    logging,
+)
+
+
+if is_transformers_available():
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+
+if is_accelerate_available():
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+TEXT_INVERSION_NAME = "learned_embeds.bin"
+TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
+
+
+def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
+    cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+    force_download = kwargs.pop("force_download", False)
+    resume_download = kwargs.pop("resume_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)
+    subfolder = kwargs.pop("subfolder", None)
+    weight_name = kwargs.pop("weight_name", None)
+    use_safetensors = kwargs.pop("use_safetensors", None)
+
+    allow_pickle = False
+    if use_safetensors is None:
+        use_safetensors = True
+        allow_pickle = True
+
+    user_agent = {
+        "file_type": "text_inversion",
+        "framework": "pytorch",
+    }
+    state_dicts = []
+    for pretrained_model_name_or_path in pretrained_model_name_or_paths:
+        if not isinstance(pretrained_model_name_or_path, (dict, torch.Tensor)):
+            # 3.1. Load textual inversion file
+            model_file = None
+
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path,
+                        weights_name=weight_name or TEXT_INVERSION_NAME_SAFE,
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        use_auth_token=use_auth_token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                    )
+                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
+                except Exception as e:
+                    if not allow_pickle:
+                        raise e
+
+                    model_file = None
+
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path,
+                    weights_name=weight_name or TEXT_INVERSION_NAME,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                state_dict = torch.load(model_file, map_location="cpu")
+        else:
+            state_dict = pretrained_model_name_or_path
+
+        state_dicts.append(state_dict)
+
+    return state_dicts
+
+
+class TextualInversionLoaderMixin:
+    r"""
+    Load textual inversion tokens and embeddings to the tokenizer and text encoder.
+    """
+
+    def maybe_convert_prompt(self, prompt: Union[str, List[str]], tokenizer: "PreTrainedTokenizer"):  # noqa: F821
+        r"""
+        Processes prompts that include a special token corresponding to a multi-vector textual inversion embedding to
+        be replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+        inversion token or if the textual inversion token is a single vector, the input prompt is returned.
+
+        Parameters:
+            prompt (`str` or list of `str`):
+                The prompt or prompts to guide the image generation.
+            tokenizer (`PreTrainedTokenizer`):
+                The tokenizer responsible for encoding the prompt into input tokens.
+
+        Returns:
+            `str` or list of `str`: The converted prompt
+        """
+        if not isinstance(prompt, List):
+            prompts = [prompt]
+        else:
+            prompts = prompt
+
+        prompts = [self._maybe_convert_prompt(p, tokenizer) for p in prompts]
+
+        if not isinstance(prompt, List):
+            return prompts[0]
+
+        return prompts
+
+    def _maybe_convert_prompt(self, prompt: str, tokenizer: "PreTrainedTokenizer"):  # noqa: F821
+        r"""
+        Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
+        to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
+        is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+        inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
+
+        Parameters:
+            prompt (`str`):
+                The prompt to guide the image generation.
+            tokenizer (`PreTrainedTokenizer`):
+                The tokenizer responsible for encoding the prompt into input tokens.
+
+        Returns:
+            `str`: The converted prompt
+        """
+        tokens = tokenizer.tokenize(prompt)
+        unique_tokens = set(tokens)
+        for token in unique_tokens:
+            if token in tokenizer.added_tokens_encoder:
+                replacement = token
+                i = 1
+                while f"{token}_{i}" in tokenizer.added_tokens_encoder:
+                    replacement += f" {token}_{i}"
+                    i += 1
+
+                prompt = prompt.replace(token, replacement)
+
+        return prompt
+
+    def _check_text_inv_inputs(self, tokenizer, text_encoder, pretrained_model_name_or_paths, tokens):
+        if tokenizer is None:
+            raise ValueError(
+                f"{self.__class__.__name__} requires `self.tokenizer` or passing a `tokenizer` of type `PreTrainedTokenizer` for calling"
+                f" `{self.load_textual_inversion.__name__}`"
+            )
+
+        if text_encoder is None:
+            raise ValueError(
+                f"{self.__class__.__name__} requires `self.text_encoder` or passing a `text_encoder` of type `PreTrainedModel` for calling"
+                f" `{self.load_textual_inversion.__name__}`"
+            )
+
+        if len(pretrained_model_name_or_paths) != len(tokens):
+            raise ValueError(
+                f"You have passed a list of models of length {len(pretrained_model_name_or_paths)}, and list of tokens of length {len(tokens)} "
+                f"Make sure both lists have the same length."
+            )
+
+        valid_tokens = [t for t in tokens if t is not None]
+        if len(set(valid_tokens)) < len(valid_tokens):
+            raise ValueError(f"You have passed a list of tokens that contains duplicates: {tokens}")
+
+    @staticmethod
+    def _retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer):
+        all_tokens = []
+        all_embeddings = []
+        for state_dict, token in zip(state_dicts, tokens):
+            if isinstance(state_dict, torch.Tensor):
+                if token is None:
+                    raise ValueError(
+                        "You are trying to load a textual inversion embedding that has been saved as a PyTorch tensor. Make sure to pass the name of the corresponding token in this case: `token=...`."
+                    )
+                loaded_token = token
+                embedding = state_dict
+            elif len(state_dict) == 1:
+                # diffusers
+                loaded_token, embedding = next(iter(state_dict.items()))
+            elif "string_to_param" in state_dict:
+                # A1111
+                loaded_token = state_dict["name"]
+                embedding = state_dict["string_to_param"]["*"]
+            else:
+                raise ValueError(
+                    f"Loaded state dictonary is incorrect: {state_dict}. \n\n"
+                    "Please verify that the loaded state dictionary of the textual embedding either only has a single key or includes the `string_to_param`"
+                    " input key."
+                )
+
+            if token is not None and loaded_token != token:
+                logger.info(f"The loaded token: {loaded_token} is overwritten by the passed token {token}.")
+            else:
+                token = loaded_token
+
+            if token in tokenizer.get_vocab():
+                raise ValueError(
+                    f"Token {token} already in tokenizer vocabulary. Please choose a different token name or remove {token} and embedding from the tokenizer and text encoder."
+                )
+
+            all_tokens.append(token)
+            all_embeddings.append(embedding)
+
+        return all_tokens, all_embeddings
+
+    @staticmethod
+    def _extend_tokens_and_embeddings(tokens, embeddings, tokenizer):
+        all_tokens = []
+        all_embeddings = []
+
+        for embedding, token in zip(embeddings, tokens):
+            if f"{token}_1" in tokenizer.get_vocab():
+                multi_vector_tokens = [token]
+                i = 1
+                while f"{token}_{i}" in tokenizer.added_tokens_encoder:
+                    multi_vector_tokens.append(f"{token}_{i}")
+                    i += 1
+
+                raise ValueError(
+                    f"Multi-vector Token {multi_vector_tokens} already in tokenizer vocabulary. Please choose a different token name or remove the {multi_vector_tokens} and embedding from the tokenizer and text encoder."
+                )
+
+            is_multi_vector = len(embedding.shape) > 1 and embedding.shape[0] > 1
+            if is_multi_vector:
+                all_tokens += [token] + [f"{token}_{i}" for i in range(1, embedding.shape[0])]
+                all_embeddings += [e for e in embedding]  # noqa: C416
+            else:
+                all_tokens += [token]
+                all_embeddings += [embedding[0]] if len(embedding.shape) > 1 else [embedding]
+
+        return all_tokens, all_embeddings
+
+    def load_textual_inversion(
+        self,
+        pretrained_model_name_or_path: Union[str, List[str], Dict[str, torch.Tensor], List[Dict[str, torch.Tensor]]],
+        token: Optional[Union[str, List[str]]] = None,
+        tokenizer: Optional["PreTrainedTokenizer"] = None,  # noqa: F821
+        text_encoder: Optional["PreTrainedModel"] = None,  # noqa: F821
+        **kwargs,
+    ):
+        r"""
+        Load textual inversion embeddings into the text encoder of [`StableDiffusionPipeline`] (both 🤗 Diffusers and
+        Automatic1111 formats are supported).
+
+        Parameters:
+            pretrained_model_name_or_path (`str` or `os.PathLike` or `List[str or os.PathLike]` or `Dict` or `List[Dict]`):
+                Can be either one of the following or a list of them:
+
+                    - A string, the *model id* (for example `sd-concepts-library/low-poly-hd-logos-icons`) of a
+                      pretrained model hosted on the Hub.
+                    - A path to a *directory* (for example `./my_text_inversion_directory/`) containing the textual
+                      inversion weights.
+                    - A path to a *file* (for example `./my_text_inversions.pt`) containing textual inversion weights.
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            token (`str` or `List[str]`, *optional*):
+                Override the token to use for the textual inversion weights. If `pretrained_model_name_or_path` is a
+                list, then `token` must also be a list of equal length.
+            text_encoder ([`~transformers.CLIPTextModel`], *optional*):
+                Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+                If not specified, function will take self.tokenizer.
+            tokenizer ([`~transformers.CLIPTokenizer`], *optional*):
+                A `CLIPTokenizer` to tokenize text. If not specified, function will take self.tokenizer.
+            weight_name (`str`, *optional*):
+                Name of a custom weight file. This should be used when:
+
+                    - The saved textual inversion file is in 🤗 Diffusers format, but was saved under a specific weight
+                      name such as `text_inv.bin`.
+                    - The saved textual inversion file is in the Automatic1111 format.
+            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.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                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.
+            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.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            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.
+
+        Example:
+
+        To load a textual inversion embedding vector in 🤗 Diffusers format:
+
+        ```py
+        from diffusers import StableDiffusionPipeline
+        import torch
+
+        model_id = "runwayml/stable-diffusion-v1-5"
+        pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
+
+        pipe.load_textual_inversion("sd-concepts-library/cat-toy")
+
+        prompt = "A <cat-toy> backpack"
+
+        image = pipe(prompt, num_inference_steps=50).images[0]
+        image.save("cat-backpack.png")
+        ```
+
+        To load a textual inversion embedding vector in Automatic1111 format, make sure to download the vector first
+        (for example from [civitAI](https://civitai.com/models/3036?modelVersionId=9857)) and then load the vector
+        locally:
+
+        ```py
+        from diffusers import StableDiffusionPipeline
+        import torch
+
+        model_id = "runwayml/stable-diffusion-v1-5"
+        pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
+
+        pipe.load_textual_inversion("./charturnerv2.pt", token="charturnerv2")
+
+        prompt = "charturnerv2, multiple views of the same character in the same outfit, a character turnaround of a woman wearing a black jacket and red shirt, best quality, intricate details."
+
+        image = pipe(prompt, num_inference_steps=50).images[0]
+        image.save("character.png")
+        ```
+
+        """
+        # 1. Set correct tokenizer and text encoder
+        tokenizer = tokenizer or getattr(self, "tokenizer", None)
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+
+        # 2. Normalize inputs
+        pretrained_model_name_or_paths = (
+            [pretrained_model_name_or_path]
+            if not isinstance(pretrained_model_name_or_path, list)
+            else pretrained_model_name_or_path
+        )
+        tokens = len(pretrained_model_name_or_paths) * [token] if (isinstance(token, str) or token is None) else token
+
+        # 3. Check inputs
+        self._check_text_inv_inputs(tokenizer, text_encoder, pretrained_model_name_or_paths, tokens)
+
+        # 4. Load state dicts of textual embeddings
+        state_dicts = load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
+
+        # 4. Retrieve tokens and embeddings
+        tokens, embeddings = self._retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer)
+
+        # 5. Extend tokens and embeddings for multi vector
+        tokens, embeddings = self._extend_tokens_and_embeddings(tokens, embeddings, tokenizer)
+
+        # 6. Make sure all embeddings have the correct size
+        expected_emb_dim = text_encoder.get_input_embeddings().weight.shape[-1]
+        if any(expected_emb_dim != emb.shape[-1] for emb in embeddings):
+            raise ValueError(
+                "Loaded embeddings are of incorrect shape. Expected each textual inversion embedding "
+                "to be of shape {input_embeddings.shape[-1]}, but are {embeddings.shape[-1]} "
+            )
+
+        # 7. Now we can be sure that loading the embedding matrix works
+        # < Unsafe code:
+
+        # 7.1 Offload all hooks in case the pipeline was cpu offloaded before make sure, we offload and onload again
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        for _, component in self.components.items():
+            if isinstance(component, nn.Module):
+                if hasattr(component, "_hf_hook"):
+                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    logger.info(
+                        "Accelerate hooks detected. Since you have called `load_textual_inversion()`, the previous hooks will be first removed. Then the textual inversion parameters will be loaded and the hooks will be applied again."
+                    )
+                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+        # 7.2 save expected device and dtype
+        device = text_encoder.device
+        dtype = text_encoder.dtype
+
+        # 7.3 Increase token embedding matrix
+        text_encoder.resize_token_embeddings(len(tokenizer) + len(tokens))
+        input_embeddings = text_encoder.get_input_embeddings().weight
+
+        # 7.4 Load token and embedding
+        for token, embedding in zip(tokens, embeddings):
+            # add tokens and get ids
+            tokenizer.add_tokens(token)
+            token_id = tokenizer.convert_tokens_to_ids(token)
+            input_embeddings.data[token_id] = embedding
+            logger.info(f"Loaded textual inversion embedding for {token}.")
+
+        input_embeddings.to(dtype=dtype, device=device)
+
+        # 7.5 Offload the model again
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
+        # / Unsafe Code >

src/diffusers/loaders/unet.py

+# Copyright 2023 The HuggingFace Team. 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 os
+from collections import defaultdict
+from contextlib import nullcontext
+from typing import Callable, Dict, List, Optional, Union
+
+import safetensors
+import torch
+from torch import nn
+
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+from ..utils import (
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
+    USE_PEFT_BACKEND,
+    _get_model_file,
+    delete_adapter_layers,
+    is_accelerate_available,
+    logging,
+    set_adapter_layers,
+    set_weights_and_activate_adapters,
+)
+from .utils import AttnProcsLayers
+
+
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
+
+LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
+CUSTOM_DIFFUSION_WEIGHT_NAME = "pytorch_custom_diffusion_weights.bin"
+CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE = "pytorch_custom_diffusion_weights.safetensors"
+
+
+class UNet2DConditionLoadersMixin:
+    text_encoder_name = TEXT_ENCODER_NAME
+    unet_name = UNET_NAME
+
+    def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
+        r"""
+        Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
+        defined in
+        [`attention_processor.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py)
+        and be a `torch.nn.Module` class.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the model id (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a directory (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            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.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                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.
+            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.
+            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.
+            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.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            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.
+
+        """
+        from ..models.attention_processor import CustomDiffusionAttnProcessor
+        from ..models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
+
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_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)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
+        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+        network_alphas = kwargs.pop("network_alphas", None)
+
+        _pipeline = kwargs.pop("_pipeline", None)
+
+        is_network_alphas_none = network_alphas is None
+
+        allow_pickle = False
+
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+
+        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."
+            )
+
+        model_file = None
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path_or_dict,
+                        weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        use_auth_token=use_auth_token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                    )
+                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
+                except IOError as e:
+                    if not allow_pickle:
+                        raise e
+                    # try loading non-safetensors weights
+                    pass
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=weight_name or LORA_WEIGHT_NAME,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                state_dict = torch.load(model_file, map_location="cpu")
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        # fill attn processors
+        lora_layers_list = []
+
+        is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys()) and not USE_PEFT_BACKEND
+        is_custom_diffusion = any("custom_diffusion" in k for k in state_dict.keys())
+
+        if is_lora:
+            # correct keys
+            state_dict, network_alphas = self.convert_state_dict_legacy_attn_format(state_dict, network_alphas)
+
+            if network_alphas is not None:
+                network_alphas_keys = list(network_alphas.keys())
+                used_network_alphas_keys = set()
+
+            lora_grouped_dict = defaultdict(dict)
+            mapped_network_alphas = {}
+
+            all_keys = list(state_dict.keys())
+            for key in all_keys:
+                value = state_dict.pop(key)
+                attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                lora_grouped_dict[attn_processor_key][sub_key] = value
+
+                # Create another `mapped_network_alphas` dictionary so that we can properly map them.
+                if network_alphas is not None:
+                    for k in network_alphas_keys:
+                        if k.replace(".alpha", "") in key:
+                            mapped_network_alphas.update({attn_processor_key: network_alphas.get(k)})
+                            used_network_alphas_keys.add(k)
+
+            if not is_network_alphas_none:
+                if len(set(network_alphas_keys) - used_network_alphas_keys) > 0:
+                    raise ValueError(
+                        f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
+                    )
+
+            if len(state_dict) > 0:
+                raise ValueError(
+                    f"The `state_dict` has to be empty at this point but has the following keys \n\n {', '.join(state_dict.keys())}"
+                )
+
+            for key, value_dict in lora_grouped_dict.items():
+                attn_processor = self
+                for sub_key in key.split("."):
+                    attn_processor = getattr(attn_processor, sub_key)
+
+                # Process non-attention layers, which don't have to_{k,v,q,out_proj}_lora layers
+                # or add_{k,v,q,out_proj}_proj_lora layers.
+                rank = value_dict["lora.down.weight"].shape[0]
+
+                if isinstance(attn_processor, LoRACompatibleConv):
+                    in_features = attn_processor.in_channels
+                    out_features = attn_processor.out_channels
+                    kernel_size = attn_processor.kernel_size
+
+                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+                    with ctx():
+                        lora = LoRAConv2dLayer(
+                            in_features=in_features,
+                            out_features=out_features,
+                            rank=rank,
+                            kernel_size=kernel_size,
+                            stride=attn_processor.stride,
+                            padding=attn_processor.padding,
+                            network_alpha=mapped_network_alphas.get(key),
+                        )
+                elif isinstance(attn_processor, LoRACompatibleLinear):
+                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+                    with ctx():
+                        lora = LoRALinearLayer(
+                            attn_processor.in_features,
+                            attn_processor.out_features,
+                            rank,
+                            mapped_network_alphas.get(key),
+                        )
+                else:
+                    raise ValueError(f"Module {key} is not a LoRACompatibleConv or LoRACompatibleLinear module.")
+
+                value_dict = {k.replace("lora.", ""): v for k, v in value_dict.items()}
+                lora_layers_list.append((attn_processor, lora))
+
+                if low_cpu_mem_usage:
+                    device = next(iter(value_dict.values())).device
+                    dtype = next(iter(value_dict.values())).dtype
+                    load_model_dict_into_meta(lora, value_dict, device=device, dtype=dtype)
+                else:
+                    lora.load_state_dict(value_dict)
+
+        elif is_custom_diffusion:
+            attn_processors = {}
+            custom_diffusion_grouped_dict = defaultdict(dict)
+            for key, value in state_dict.items():
+                if len(value) == 0:
+                    custom_diffusion_grouped_dict[key] = {}
+                else:
+                    if "to_out" in key:
+                        attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                    else:
+                        attn_processor_key, sub_key = ".".join(key.split(".")[:-2]), ".".join(key.split(".")[-2:])
+                    custom_diffusion_grouped_dict[attn_processor_key][sub_key] = value
+
+            for key, value_dict in custom_diffusion_grouped_dict.items():
+                if len(value_dict) == 0:
+                    attn_processors[key] = CustomDiffusionAttnProcessor(
+                        train_kv=False, train_q_out=False, hidden_size=None, cross_attention_dim=None
+                    )
+                else:
+                    cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[1]
+                    hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[0]
+                    train_q_out = True if "to_q_custom_diffusion.weight" in value_dict else False
+                    attn_processors[key] = CustomDiffusionAttnProcessor(
+                        train_kv=True,
+                        train_q_out=train_q_out,
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                    )
+                    attn_processors[key].load_state_dict(value_dict)
+        elif USE_PEFT_BACKEND:
+            # In that case we have nothing to do as loading the adapter weights is already handled above by `set_peft_model_state_dict`
+            # on the Unet
+            pass
+        else:
+            raise ValueError(
+                f"{model_file} does not seem to be in the correct format expected by LoRA or Custom Diffusion training."
+            )
+
+        # <Unsafe code
+        # We can be sure that the following works as it just sets attention processors, lora layers and puts all in the same dtype
+        # Now we remove any existing hooks to
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+
+        # For PEFT backend the Unet is already offloaded at this stage as it is handled inside `lora_lora_weights_into_unet`
+        if not USE_PEFT_BACKEND:
+            if _pipeline is not None:
+                for _, component in _pipeline.components.items():
+                    if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
+                        is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                        is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+
+                        logger.info(
+                            "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                        )
+                        remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+            # only custom diffusion needs to set attn processors
+            if is_custom_diffusion:
+                self.set_attn_processor(attn_processors)
+
+            # set lora layers
+            for target_module, lora_layer in lora_layers_list:
+                target_module.set_lora_layer(lora_layer)
+
+            self.to(dtype=self.dtype, device=self.device)
+
+            # Offload back.
+            if is_model_cpu_offload:
+                _pipeline.enable_model_cpu_offload()
+            elif is_sequential_cpu_offload:
+                _pipeline.enable_sequential_cpu_offload()
+            # Unsafe code />
+
+    def convert_state_dict_legacy_attn_format(self, state_dict, network_alphas):
+        is_new_lora_format = all(
+            key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
+        )
+        if is_new_lora_format:
+            # Strip the `"unet"` prefix.
+            is_text_encoder_present = any(key.startswith(self.text_encoder_name) for key in state_dict.keys())
+            if is_text_encoder_present:
+                warn_message = "The state_dict contains LoRA params corresponding to the text encoder which are not being used here. To use both UNet and text encoder related LoRA params, use [`pipe.load_lora_weights()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights)."
+                logger.warn(warn_message)
+            unet_keys = [k for k in state_dict.keys() if k.startswith(self.unet_name)]
+            state_dict = {k.replace(f"{self.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
+
+        # change processor format to 'pure' LoRACompatibleLinear format
+        if any("processor" in k.split(".") for k in state_dict.keys()):
+
+            def format_to_lora_compatible(key):
+                if "processor" not in key.split("."):
+                    return key
+                return key.replace(".processor", "").replace("to_out_lora", "to_out.0.lora").replace("_lora", ".lora")
+
+            state_dict = {format_to_lora_compatible(k): v for k, v in state_dict.items()}
+
+            if network_alphas is not None:
+                network_alphas = {format_to_lora_compatible(k): v for k, v in network_alphas.items()}
+        return state_dict, network_alphas
+
+    def save_attn_procs(
+        self,
+        save_directory: Union[str, os.PathLike],
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        **kwargs,
+    ):
+        r"""
+        Save an attention processor to a directory so that it can be reloaded using the
+        [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] method.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save an attention processor to. Will be created if it doesn't exist.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+        """
+        from ..models.attention_processor import (
+            CustomDiffusionAttnProcessor,
+            CustomDiffusionAttnProcessor2_0,
+            CustomDiffusionXFormersAttnProcessor,
+        )
+
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        if save_function is None:
+            if safe_serialization:
+
+                def save_function(weights, filename):
+                    return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
+
+            else:
+                save_function = torch.save
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        is_custom_diffusion = any(
+            isinstance(
+                x,
+                (CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor),
+            )
+            for (_, x) in self.attn_processors.items()
+        )
+        if is_custom_diffusion:
+            model_to_save = AttnProcsLayers(
+                {
+                    y: x
+                    for (y, x) in self.attn_processors.items()
+                    if isinstance(
+                        x,
+                        (
+                            CustomDiffusionAttnProcessor,
+                            CustomDiffusionAttnProcessor2_0,
+                            CustomDiffusionXFormersAttnProcessor,
+                        ),
+                    )
+                }
+            )
+            state_dict = model_to_save.state_dict()
+            for name, attn in self.attn_processors.items():
+                if len(attn.state_dict()) == 0:
+                    state_dict[name] = {}
+        else:
+            model_to_save = AttnProcsLayers(self.attn_processors)
+            state_dict = model_to_save.state_dict()
+
+        if weight_name is None:
+            if safe_serialization:
+                weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE if is_custom_diffusion else LORA_WEIGHT_NAME_SAFE
+            else:
+                weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME if is_custom_diffusion else LORA_WEIGHT_NAME
+
+        # Save the model
+        save_function(state_dict, os.path.join(save_directory, weight_name))
+        logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
+
+    def fuse_lora(self, lora_scale=1.0, safe_fusing=False):
+        self.lora_scale = lora_scale
+        self._safe_fusing = safe_fusing
+        self.apply(self._fuse_lora_apply)
+
+    def _fuse_lora_apply(self, module):
+        if not USE_PEFT_BACKEND:
+            if hasattr(module, "_fuse_lora"):
+                module._fuse_lora(self.lora_scale, self._safe_fusing)
+        else:
+            from peft.tuners.tuners_utils import BaseTunerLayer
+
+            if isinstance(module, BaseTunerLayer):
+                if self.lora_scale != 1.0:
+                    module.scale_layer(self.lora_scale)
+                module.merge(safe_merge=self._safe_fusing)
+
+    def unfuse_lora(self):
+        self.apply(self._unfuse_lora_apply)
+
+    def _unfuse_lora_apply(self, module):
+        if not USE_PEFT_BACKEND:
+            if hasattr(module, "_unfuse_lora"):
+                module._unfuse_lora()
+        else:
+            from peft.tuners.tuners_utils import BaseTunerLayer
+
+            if isinstance(module, BaseTunerLayer):
+                module.unmerge()
+
+    def set_adapters(
+        self,
+        adapter_names: Union[List[str], str],
+        weights: Optional[Union[List[float], float]] = None,
+    ):
+        """
+        Sets the adapter layers for the unet.
+
+        Args:
+            adapter_names (`List[str]` or `str`):
+                The names of the adapters to use.
+            weights (`Union[List[float], float]`, *optional*):
+                The adapter(s) weights to use with the UNet. If `None`, the weights are set to `1.0` for all the
+                adapters.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for `set_adapters()`.")
+
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+
+        if weights is None:
+            weights = [1.0] * len(adapter_names)
+        elif isinstance(weights, float):
+            weights = [weights] * len(adapter_names)
+
+        if len(adapter_names) != len(weights):
+            raise ValueError(
+                f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
+            )
+
+        set_weights_and_activate_adapters(self, adapter_names, weights)
+
+    def disable_lora(self):
+        """
+        Disables the active LoRA layers for the unet.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        set_adapter_layers(self, enabled=False)
+
+    def enable_lora(self):
+        """
+        Enables the active LoRA layers for the unet.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        set_adapter_layers(self, enabled=True)
+
+    def delete_adapters(self, adapter_names: Union[List[str], str]):
+        """
+        Args:
+        Deletes the LoRA layers of `adapter_name` for the unet.
+            adapter_names (`Union[List[str], str]`):
+                The names of the adapter to delete. Can be a single string or a list of strings
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        if isinstance(adapter_names, str):
+            adapter_names = [adapter_names]
+
+        for adapter_name in adapter_names:
+            delete_adapter_layers(self, adapter_name)
+
+            # Pop also the corresponding adapter from the config
+            if hasattr(self, "peft_config"):
+                self.peft_config.pop(adapter_name, None)
+
+    delete_adapter_layers

src/diffusers/loaders/utils.py

+# Copyright 2023 The HuggingFace Team. 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.
+
+from typing import Dict
+
+import torch
+
+
+class AttnProcsLayers(torch.nn.Module):
+    def __init__(self, state_dict: Dict[str, torch.Tensor]):
+        super().__init__()
+        self.layers = torch.nn.ModuleList(state_dict.values())
+        self.mapping = dict(enumerate(state_dict.keys()))
+        self.rev_mapping = {v: k for k, v in enumerate(state_dict.keys())}
+
+        # .processor for unet, .self_attn for text encoder
+        self.split_keys = [".processor", ".self_attn"]
+
+        # we add a hook to state_dict() and load_state_dict() so that the
+        # naming fits with `unet.attn_processors`
+        def map_to(module, state_dict, *args, **kwargs):
+            new_state_dict = {}
+            for key, value in state_dict.items():
+                num = int(key.split(".")[1])  # 0 is always "layers"
+                new_key = key.replace(f"layers.{num}", module.mapping[num])
+                new_state_dict[new_key] = value
+
+            return new_state_dict
+
+        def remap_key(key, state_dict):
+            for k in self.split_keys:
+                if k in key:
+                    return key.split(k)[0] + k
+
+            raise ValueError(
+                f"There seems to be a problem with the state_dict: {set(state_dict.keys())}. {key} has to have one of {self.split_keys}."
+            )
+
+        def map_from(module, state_dict, *args, **kwargs):
+            all_keys = list(state_dict.keys())
+            for key in all_keys:
+                replace_key = remap_key(key, state_dict)
+                new_key = key.replace(replace_key, f"layers.{module.rev_mapping[replace_key]}")
+                state_dict[new_key] = state_dict[key]
+                del state_dict[key]
+
+        self._register_state_dict_hook(map_to)
+        self._register_load_state_dict_pre_hook(map_from, with_module=True)

src/diffusers/models/lora.py

@@ -18,13 +18,64 @@ import torch
 import torch.nn.functional as F
 from torch import nn
 
-from ..loaders import PatchedLoraProjection, text_encoder_attn_modules, text_encoder_mlp_modules
 from ..utils import logging
+from ..utils.import_utils import is_transformers_available
+
+
+if is_transformers_available():
+    from transformers import CLIPTextModel, CLIPTextModelWithProjection
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
+def text_encoder_attn_modules(text_encoder):
+    attn_modules = []
+
+    if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+        for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+            name = f"text_model.encoder.layers.{i}.self_attn"
+            mod = layer.self_attn
+            attn_modules.append((name, mod))
+    else:
+        raise ValueError(f"do not know how to get attention modules for: {text_encoder.__class__.__name__}")
+
+    return attn_modules
+
+
+def text_encoder_mlp_modules(text_encoder):
+    mlp_modules = []
+
+    if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+        for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+            mlp_mod = layer.mlp
+            name = f"text_model.encoder.layers.{i}.mlp"
+            mlp_modules.append((name, mlp_mod))
+    else:
+        raise ValueError(f"do not know how to get mlp modules for: {text_encoder.__class__.__name__}")
+
+    return mlp_modules
+
+
+def text_encoder_lora_state_dict(text_encoder):
+    state_dict = {}
+
+    for name, module in text_encoder_attn_modules(text_encoder):
+        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
+
+    return state_dict
+
+
 def adjust_lora_scale_text_encoder(text_encoder, lora_scale: float = 1.0):
     for _, attn_module in text_encoder_attn_modules(text_encoder):
         if isinstance(attn_module.q_proj, PatchedLoraProjection):
@@ -39,6 +90,95 @@ def adjust_lora_scale_text_encoder(text_encoder, lora_scale: float = 1.0):
             mlp_module.fc2.lora_scale = lora_scale
 
 
+class PatchedLoraProjection(torch.nn.Module):
+    def __init__(self, regular_linear_layer, lora_scale=1, network_alpha=None, rank=4, dtype=None):
+        super().__init__()
+        from ..models.lora import LoRALinearLayer
+
+        self.regular_linear_layer = regular_linear_layer
+
+        device = self.regular_linear_layer.weight.device
+
+        if dtype is None:
+            dtype = self.regular_linear_layer.weight.dtype
+
+        self.lora_linear_layer = LoRALinearLayer(
+            self.regular_linear_layer.in_features,
+            self.regular_linear_layer.out_features,
+            network_alpha=network_alpha,
+            device=device,
+            dtype=dtype,
+            rank=rank,
+        )
+
+        self.lora_scale = lora_scale
+
+    # overwrite PyTorch's `state_dict` to be sure that only the 'regular_linear_layer' weights are saved
+    # when saving the whole text encoder model and when LoRA is unloaded or fused
+    def state_dict(self, *args, destination=None, prefix="", keep_vars=False):
+        if self.lora_linear_layer is None:
+            return self.regular_linear_layer.state_dict(
+                *args, destination=destination, prefix=prefix, keep_vars=keep_vars
+            )
+
+        return super().state_dict(*args, destination=destination, prefix=prefix, keep_vars=keep_vars)
+
+    def _fuse_lora(self, lora_scale=1.0, safe_fusing=False):
+        if self.lora_linear_layer is None:
+            return
+
+        dtype, device = self.regular_linear_layer.weight.data.dtype, self.regular_linear_layer.weight.data.device
+
+        w_orig = self.regular_linear_layer.weight.data.float()
+        w_up = self.lora_linear_layer.up.weight.data.float()
+        w_down = self.lora_linear_layer.down.weight.data.float()
+
+        if self.lora_linear_layer.network_alpha is not None:
+            w_up = w_up * self.lora_linear_layer.network_alpha / self.lora_linear_layer.rank
+
+        fused_weight = w_orig + (lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
+
+        if safe_fusing and torch.isnan(fused_weight).any().item():
+            raise ValueError(
+                "This LoRA weight seems to be broken. "
+                f"Encountered NaN values when trying to fuse LoRA weights for {self}."
+                "LoRA weights will not be fused."
+            )
+
+        self.regular_linear_layer.weight.data = fused_weight.to(device=device, dtype=dtype)
+
+        # we can drop the lora layer now
+        self.lora_linear_layer = None
+
+        # offload the up and down matrices to CPU to not blow the memory
+        self.w_up = w_up.cpu()
+        self.w_down = w_down.cpu()
+        self.lora_scale = lora_scale
+
+    def _unfuse_lora(self):
+        if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
+            return
+
+        fused_weight = self.regular_linear_layer.weight.data
+        dtype, device = fused_weight.dtype, fused_weight.device
+
+        w_up = self.w_up.to(device=device).float()
+        w_down = self.w_down.to(device).float()
+
+        unfused_weight = fused_weight.float() - (self.lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
+        self.regular_linear_layer.weight.data = unfused_weight.to(device=device, dtype=dtype)
+
+        self.w_up = None
+        self.w_down = None
+
+    def forward(self, input):
+        if self.lora_scale is None:
+            self.lora_scale = 1.0
+        if self.lora_linear_layer is None:
+            return self.regular_linear_layer(input)
+        return self.regular_linear_layer(input) + (self.lora_scale * self.lora_linear_layer(input))
+
+
 class LoRALinearLayer(nn.Module):
     r"""
     A linear layer that is used with LoRA.

tests/lora/test_lora_layers_old_backend.py

@@ -41,7 +41,7 @@ from diffusers import (
     UNet2DConditionModel,
     UNet3DConditionModel,
 )
-from diffusers.loaders import AttnProcsLayers, LoraLoaderMixin, PatchedLoraProjection, text_encoder_attn_modules
+from diffusers.loaders import AttnProcsLayers, LoraLoaderMixin
 from diffusers.models.attention_processor import (
     Attention,
     AttnProcessor,
@@ -51,6 +51,7 @@ from diffusers.models.attention_processor import (
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
 )
+from diffusers.models.lora import PatchedLoraProjection, text_encoder_attn_modules
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import (
     deprecate_after_peft_backend,

tests/lora/test_lora_layers_peft.py

@@ -40,10 +40,7 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.loaders import AttnProcsLayers
-from diffusers.models.attention_processor import (
-    LoRAAttnProcessor,
-    LoRAAttnProcessor2_0,
-)
+from diffusers.models.attention_processor import LoRAAttnProcessor, LoRAAttnProcessor2_0
 from diffusers.utils.import_utils import is_accelerate_available, is_peft_available
 from diffusers.utils.testing_utils import (
     floats_tensor,