diffusers 0.33.0__py3-none-any.whl → 0.34.0__py3-none-any.whl

diffusers/pipelines/paint_by_example/pipeline_paint_by_example.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -25,7 +25,7 @@ from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
 from ...utils import deprecate, is_torch_xla_available, logging
 from ...utils.torch_utils import randn_tensor
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from .image_encoder import PaintByExampleImageEncoder
@@ -155,7 +155,8 @@ def prepare_mask_and_masked_image(image, mask):
     return mask, masked_image
 
 
-class PaintByExamplePipeline(DiffusionPipeline, StableDiffusionMixin):
+class PaintByExamplePipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin):
+    _last_supported_version = "0.33.1"
     r"""
     <Tip warning={true}>
 
@@ -239,7 +240,7 @@ class PaintByExamplePipeline(DiffusionPipeline, StableDiffusionMixin):
     def prepare_extra_step_kwargs(self, generator, eta):
         # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
         # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
-        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502
         # and should be between [0, 1]
 
         accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
@@ -447,8 +448,8 @@ class PaintByExamplePipeline(DiffusionPipeline, StableDiffusionMixin):
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             eta (`float`, *optional*, defaults to 0.0):
-                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
-                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+                Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only
+                applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                 A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                 generation deterministic.
@@ -521,7 +522,7 @@ class PaintByExamplePipeline(DiffusionPipeline, StableDiffusionMixin):
             batch_size = image.shape[0]
         device = self._execution_device
         # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
-        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
         # corresponds to doing no classifier free guidance.
         do_classifier_free_guidance = guidance_scale > 1.0
 

diffusers/pipelines/pia/pipeline_pia.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -46,7 +46,7 @@ from ...utils import (
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
 from ..free_init_utils import FreeInitMixin
-from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 
 
 if is_torch_xla_available():
@@ -132,6 +132,7 @@ class PIAPipelineOutput(BaseOutput):
 
 
 class PIAPipeline(
+    DeprecatedPipelineMixin,
     DiffusionPipeline,
     StableDiffusionMixin,
     TextualInversionLoaderMixin,
@@ -140,6 +141,7 @@ class PIAPipeline(
     FromSingleFileMixin,
     FreeInitMixin,
 ):
+    _last_supported_version = "0.33.1"
     r"""
     Pipeline for text-to-video generation.
 
@@ -432,7 +434,7 @@ class PIAPipeline(
     def prepare_extra_step_kwargs(self, generator, eta):
         # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
         # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
-        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502
         # and should be between [0, 1]
 
         accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
@@ -653,7 +655,7 @@ class PIAPipeline(
         return self._clip_skip
 
     # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
-    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
@@ -723,8 +725,8 @@ class PIAPipeline(
                 The prompt or prompts to guide what to not include in image generation. If not defined, you need to
                 pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
             eta (`float`, *optional*, defaults to 0.0):
-                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
-                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+                Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only
+                applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                 A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                 generation deterministic.

diffusers/pipelines/pipeline_flax_utils.py

@@ -248,9 +248,8 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
                       pretrained pipeline hosted on the Hub.
                     - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
                       using [`~FlaxDiffusionPipeline.save_pretrained`].
-            dtype (`str` or `jnp.dtype`, *optional*):
-                Override the default `jnp.dtype` and load the model under this dtype. If `"auto"`, the dtype is
-                automatically derived from the model's weights.
+            dtype (`jnp.dtype`, *optional*):
+                Override the default `jnp.dtype` and load the model under this dtype.
             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.
@@ -469,7 +468,7 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
                 class_obj = import_flax_or_no_model(pipeline_module, class_name)
 
                 importable_classes = ALL_IMPORTABLE_CLASSES
-                class_candidates = {c: class_obj for c in importable_classes.keys()}
+                class_candidates = dict.fromkeys(importable_classes.keys(), class_obj)
             else:
                 # else we just import it from the library.
                 library = importlib.import_module(library_name)

diffusers/pipelines/pipeline_loading_utils.py

@@ -92,7 +92,7 @@ for library in LOADABLE_CLASSES:
     ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])
 
 
-def is_safetensors_compatible(filenames, passed_components=None, folder_names=None) -> bool:
+def is_safetensors_compatible(filenames, passed_components=None, folder_names=None, variant=None) -> bool:
     """
     Checking for safetensors compatibility:
     - The model is safetensors compatible only if there is a safetensors file for each model component present in
@@ -103,6 +103,31 @@ def is_safetensors_compatible(filenames, passed_components=None, folder_names=No
     - For models from the transformers library, the filename changes from "pytorch_model" to "model", and the ".bin"
       extension is replaced with ".safetensors"
     """
+    weight_names = [
+        WEIGHTS_NAME,
+        SAFETENSORS_WEIGHTS_NAME,
+        FLAX_WEIGHTS_NAME,
+        ONNX_WEIGHTS_NAME,
+        ONNX_EXTERNAL_WEIGHTS_NAME,
+    ]
+
+    if is_transformers_available():
+        weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME, TRANSFORMERS_FLAX_WEIGHTS_NAME]
+
+    # model_pytorch, diffusion_model_pytorch, ...
+    weight_prefixes = [w.split(".")[0] for w in weight_names]
+    # .bin, .safetensors, ...
+    weight_suffixs = [w.split(".")[-1] for w in weight_names]
+    # -00001-of-00002
+    transformers_index_format = r"\d{5}-of-\d{5}"
+    # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors`
+    variant_file_re = re.compile(
+        rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$"
+    )
+    non_variant_file_re = re.compile(
+        rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$"
+    )
+
     passed_components = passed_components or []
     if folder_names:
         filenames = {f for f in filenames if os.path.split(f)[0] in folder_names}
@@ -121,15 +146,29 @@ def is_safetensors_compatible(filenames, passed_components=None, folder_names=No
         components[component].append(component_filename)
 
     # If there are no component folders check the main directory for safetensors files
+    filtered_filenames = set()
     if not components:
-        return any(".safetensors" in filename for filename in filenames)
+        if variant is not None:
+            filtered_filenames = filter_with_regex(filenames, variant_file_re)
+
+        # If no variant filenames exist check if non-variant files are available
+        if not filtered_filenames:
+            filtered_filenames = filter_with_regex(filenames, non_variant_file_re)
+        return any(".safetensors" in filename for filename in filtered_filenames)
 
     # iterate over all files of a component
     # check if safetensor files exist for that component
-    # if variant is provided check if the variant of the safetensors exists
     for component, component_filenames in components.items():
         matches = []
-        for component_filename in component_filenames:
+        filtered_component_filenames = set()
+        # if variant is provided check if the variant of the safetensors exists
+        if variant is not None:
+            filtered_component_filenames = filter_with_regex(component_filenames, variant_file_re)
+
+        # if variant safetensor files do not exist check for non-variants
+        if not filtered_component_filenames:
+            filtered_component_filenames = filter_with_regex(component_filenames, non_variant_file_re)
+        for component_filename in filtered_component_filenames:
             filename, extension = os.path.splitext(component_filename)
 
             match_exists = extension == ".safetensors"
@@ -159,6 +198,10 @@ def filter_model_files(filenames):
     return [f for f in filenames if any(f.endswith(extension) for extension in allowed_extensions)]
 
 
+def filter_with_regex(filenames, pattern_re):
+    return {f for f in filenames if pattern_re.match(f.split("/")[-1]) is not None}
+
+
 def variant_compatible_siblings(filenames, variant=None, ignore_patterns=None) -> Union[List[os.PathLike], str]:
     weight_names = [
         WEIGHTS_NAME,
@@ -207,9 +250,6 @@ def variant_compatible_siblings(filenames, variant=None, ignore_patterns=None) -
         # interested in the extension name
         return {f for f in filenames if not any(f.endswith(pat.lstrip("*.")) for pat in ignore_patterns)}
 
-    def filter_with_regex(filenames, pattern_re):
-        return {f for f in filenames if pattern_re.match(f.split("/")[-1]) is not None}
-
     # Group files by component
     components = {}
     for filename in filenames:
@@ -335,19 +375,19 @@ def get_class_obj_and_candidates(
     library_name, class_name, importable_classes, pipelines, is_pipeline_module, component_name=None, cache_dir=None
 ):
     """Simple helper method to retrieve class object of module as well as potential parent class objects"""
-    component_folder = os.path.join(cache_dir, component_name)
+    component_folder = os.path.join(cache_dir, component_name) if component_name and cache_dir else None
 
     if is_pipeline_module:
         pipeline_module = getattr(pipelines, library_name)
 
         class_obj = getattr(pipeline_module, class_name)
-        class_candidates = {c: class_obj for c in importable_classes.keys()}
-    elif os.path.isfile(os.path.join(component_folder, library_name + ".py")):
+        class_candidates = dict.fromkeys(importable_classes.keys(), class_obj)
+    elif component_folder and os.path.isfile(os.path.join(component_folder, library_name + ".py")):
         # load custom component
         class_obj = get_class_from_dynamic_module(
             component_folder, module_file=library_name + ".py", class_name=class_name
         )
-        class_candidates = {c: class_obj for c in importable_classes.keys()}
+        class_candidates = dict.fromkeys(importable_classes.keys(), class_obj)
     else:
         # else we just import it from the library.
         library = importlib.import_module(library_name)
@@ -675,8 +715,10 @@ def load_sub_model(
     use_safetensors: bool,
     dduf_entries: Optional[Dict[str, DDUFEntry]],
     provider_options: Any,
+    quantization_config: Optional[Any] = None,
 ):
     """Helper method to load the module `name` from `library_name` and `class_name`"""
+    from ..quantizers import PipelineQuantizationConfig
 
     # retrieve class candidates
 
@@ -769,6 +811,17 @@ def load_sub_model(
         else:
             loading_kwargs["low_cpu_mem_usage"] = False
 
+    if (
+        quantization_config is not None
+        and isinstance(quantization_config, PipelineQuantizationConfig)
+        and issubclass(class_obj, torch.nn.Module)
+    ):
+        model_quant_config = quantization_config._resolve_quant_config(
+            is_diffusers=is_diffusers_model, module_name=name
+        )
+        if model_quant_config is not None:
+            loading_kwargs["quantization_config"] = model_quant_config
+
     # check if the module is in a subdirectory
     if dduf_entries:
         loading_kwargs["dduf_entries"] = dduf_entries
@@ -984,7 +1037,7 @@ def _get_ignore_patterns(
         use_safetensors
         and not allow_pickle
         and not is_safetensors_compatible(
-            model_filenames, passed_components=passed_components, folder_names=model_folder_names
+            model_filenames, passed_components=passed_components, folder_names=model_folder_names, variant=variant
         )
     ):
         raise EnvironmentError(
@@ -995,7 +1048,7 @@ def _get_ignore_patterns(
         ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb"]
 
     elif use_safetensors and is_safetensors_compatible(
-        model_filenames, passed_components=passed_components, folder_names=model_folder_names
+        model_filenames, passed_components=passed_components, folder_names=model_folder_names, variant=variant
     ):
         ignore_patterns = ["*.bin", "*.msgpack"]
 
@@ -1078,3 +1131,26 @@ def _maybe_raise_error_for_incorrect_transformers(config_dict):
                 break
     if has_transformers_component and not is_transformers_version(">", "4.47.1"):
         raise ValueError("Please upgrade your `transformers` installation to the latest version to use DDUF.")
+
+
+def _maybe_warn_for_wrong_component_in_quant_config(pipe_init_dict, quant_config):
+    if quant_config is None:
+        return
+
+    actual_pipe_components = set(pipe_init_dict.keys())
+    missing = ""
+    quant_components = None
+    if getattr(quant_config, "components_to_quantize", None) is not None:
+        quant_components = set(quant_config.components_to_quantize)
+    elif getattr(quant_config, "quant_mapping", None) is not None and isinstance(quant_config.quant_mapping, dict):
+        quant_components = set(quant_config.quant_mapping.keys())
+
+    if quant_components and not quant_components.issubset(actual_pipe_components):
+        missing = quant_components - actual_pipe_components
+
+    if missing:
+        logger.warning(
+            f"The following components in the quantization config {missing} will be ignored "
+            "as they do not belong to the underlying pipeline. Acceptable values for the pipeline "
+            f"components are: {', '.join(actual_pipe_components)}."
+        )

diffusers/pipelines/pipeline_utils.py

@@ -47,6 +47,7 @@ from ..configuration_utils import ConfigMixin
 from ..models import AutoencoderKL
 from ..models.attention_processor import FusedAttnProcessor2_0
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, ModelMixin
+from ..quantizers import PipelineQuantizationConfig
 from ..quantizers.bitsandbytes.utils import _check_bnb_status
 from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
 from ..utils import (
@@ -58,6 +59,7 @@ from ..utils import (
     _is_valid_type,
     is_accelerate_available,
     is_accelerate_version,
+    is_hpu_available,
     is_torch_npu_available,
     is_torch_version,
     is_transformers_version,
@@ -65,7 +67,7 @@ from ..utils import (
     numpy_to_pil,
 )
 from ..utils.hub_utils import _check_legacy_sharding_variant_format, load_or_create_model_card, populate_model_card
-from ..utils.torch_utils import is_compiled_module
+from ..utils.torch_utils import empty_device_cache, get_device, is_compiled_module
 
 
 if is_torch_npu_available():
@@ -86,6 +88,7 @@ from .pipeline_loading_utils import (
     _identify_model_variants,
     _maybe_raise_error_for_incorrect_transformers,
     _maybe_raise_warning_for_inpainting,
+    _maybe_warn_for_wrong_component_in_quant_config,
     _resolve_custom_pipeline_and_cls,
     _unwrap_model,
     _update_init_kwargs_with_connected_pipeline,
@@ -137,6 +140,43 @@ class AudioPipelineOutput(BaseOutput):
     audios: np.ndarray
 
 
+class DeprecatedPipelineMixin:
+    """
+    A mixin that can be used to mark a pipeline as deprecated.
+
+    Pipelines inheriting from this mixin will raise a warning when instantiated, indicating that they are deprecated
+    and won't receive updates past the specified version. Tests will be skipped for pipelines that inherit from this
+    mixin.
+
+    Example usage:
+    ```python
+    class MyDeprecatedPipeline(DeprecatedPipelineMixin, DiffusionPipeline):
+        _last_supported_version = "0.20.0"
+
+        def __init__(self, *args, **kwargs):
+            super().__init__(*args, **kwargs)
+    ```
+    """
+
+    # Override this in the inheriting class to specify the last version that will support this pipeline
+    _last_supported_version = None
+
+    def __init__(self, *args, **kwargs):
+        # Get the class name for the warning message
+        class_name = self.__class__.__name__
+
+        # Get the last supported version or use the current version if not specified
+        version_info = getattr(self.__class__, "_last_supported_version", __version__)
+
+        # Raise a warning that this pipeline is deprecated
+        logger.warning(
+            f"The {class_name} has been deprecated and will not receive bug fixes or feature updates after Diffusers version {version_info}. "
+        )
+
+        # Call the parent class's __init__ method
+        super().__init__(*args, **kwargs)
+
+
 class DiffusionPipeline(ConfigMixin, PushToHubMixin):
     r"""
     Base class for all pipelines.
@@ -404,6 +444,11 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             if not is_accelerate_available() or is_accelerate_version("<", "0.14.0"):
                 return False
 
+            _, _, is_loaded_in_8bit_bnb = _check_bnb_status(module)
+
+            if is_loaded_in_8bit_bnb:
+                return False
+
             return hasattr(module, "_hf_hook") and (
                 isinstance(module._hf_hook, accelerate.hooks.AlignDevicesHook)
                 or hasattr(module._hf_hook, "hooks")
@@ -445,6 +490,20 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 f"It seems like you have activated model offloading by calling `enable_model_cpu_offload`, but are now manually moving the pipeline to GPU. It is strongly recommended against doing so as memory gains from offloading are likely to be lost. Offloading automatically takes care of moving the individual components {', '.join(self.components.keys())} to GPU when needed. To make sure offloading works as expected, you should consider moving the pipeline back to CPU: `pipeline.to('cpu')` or removing the move altogether if you use offloading."
             )
 
+        # Enable generic support for Intel Gaudi accelerator using GPU/HPU migration
+        if device_type == "hpu" and kwargs.pop("hpu_migration", True) and is_hpu_available():
+            os.environ["PT_HPU_GPU_MIGRATION"] = "1"
+            logger.debug("Environment variable set: PT_HPU_GPU_MIGRATION=1")
+
+            import habana_frameworks.torch  # noqa: F401
+
+            # HPU hardware check
+            if not (hasattr(torch, "hpu") and torch.hpu.is_available()):
+                raise ValueError("You are trying to call `.to('hpu')` but HPU device is unavailable.")
+
+            os.environ["PT_HPU_MAX_COMPOUND_OP_SIZE"] = "1"
+            logger.debug("Environment variable set: PT_HPU_MAX_COMPOUND_OP_SIZE=1")
+
         module_names, _ = self._get_signature_keys(self)
         modules = [getattr(self, n, None) for n in module_names]
         modules = [m for m in modules if isinstance(m, torch.nn.Module)]
@@ -552,12 +611,12 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                       saved using
                     [`~DiffusionPipeline.save_pretrained`].
                     - A path to a *directory* (for example `./my_pipeline_directory/`) containing a dduf file
-            torch_dtype (`str` or `torch.dtype` or `dict[str, Union[str, 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. To load submodels with different dtype pass a
-                `dict` (for example `{'transformer': torch.bfloat16, 'vae': torch.float16}`). Set the default dtype for
-                unspecified components with `default` (for example `{'transformer': torch.bfloat16, 'default':
-                torch.float16}`). If a component is not specified and no default is set, `torch.float32` is used.
+            torch_dtype (`torch.dtype` or `dict[str, Union[str, torch.dtype]]`, *optional*):
+                Override the default `torch.dtype` and load the model with another dtype. To load submodels with
+                different dtype pass a `dict` (for example `{'transformer': torch.bfloat16, 'vae': torch.float16}`).
+                Set the default dtype for unspecified components with `default` (for example `{'transformer':
+                torch.bfloat16, 'default': torch.float16}`). If a component is not specified and no default is set,
+                `torch.float32` is used.
             custom_pipeline (`str`, *optional*):
 
                 <Tip warning={true}>
@@ -611,14 +670,11 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
                 guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                 information.
-            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
-                A map that specifies where each submodule should go. It doesn’t need to be defined for each
-                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
-                same device.
-
-                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
-                more information about each option see [designing a device
-                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+            device_map (`str`, *optional*):
+                Strategy that dictates how the different components of a pipeline should be placed on available
+                devices. Currently, only "balanced" `device_map` is supported. Check out
+                [this](https://huggingface.co/docs/diffusers/main/en/tutorials/inference_with_big_models#device-placement)
+                to know more.
             max_memory (`Dict`, *optional*):
                 A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                 each GPU and the available CPU RAM if unset.
@@ -705,6 +761,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         use_safetensors = kwargs.pop("use_safetensors", None)
         use_onnx = kwargs.pop("use_onnx", None)
         load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
+        quantization_config = kwargs.pop("quantization_config", None)
 
         if torch_dtype is not None and not isinstance(torch_dtype, dict) and not isinstance(torch_dtype, torch.dtype):
             torch_dtype = torch.float32
@@ -721,6 +778,9 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 " install accelerate\n```\n."
             )
 
+        if quantization_config is not None and not isinstance(quantization_config, PipelineQuantizationConfig):
+            raise ValueError("`quantization_config` must be an instance of `PipelineQuantizationConfig`.")
+
         if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
             raise NotImplementedError(
                 "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
@@ -925,6 +985,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
 
         # 7. Load each module in the pipeline
         current_device_map = None
+        _maybe_warn_for_wrong_component_in_quant_config(init_dict, quantization_config)
         for name, (library_name, class_name) in logging.tqdm(init_dict.items(), desc="Loading pipeline components..."):
             # 7.1 device_map shenanigans
             if final_device_map is not None and len(final_device_map) > 0:
@@ -981,6 +1042,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                     use_safetensors=use_safetensors,
                     dduf_entries=dduf_entries,
                     provider_options=provider_options,
+                    quantization_config=quantization_config,
                 )
                 logger.info(
                     f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}."
@@ -1084,19 +1146,20 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 accelerate.hooks.remove_hook_from_module(model, recurse=True)
         self._all_hooks = []
 
-    def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
-        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
-        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
-        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the accelerator when its
+        `forward` method is called, and the model remains in accelerator until the next model runs. Memory savings are
+        lower than with `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution
+        of the `unet`.
 
         Arguments:
             gpu_id (`int`, *optional*):
                 The ID of the accelerator that shall be used in inference. If not specified, it will default to 0.
-            device (`torch.Device` or `str`, *optional*, defaults to "cuda"):
+            device (`torch.Device` or `str`, *optional*, defaults to None):
                 The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will
-                default to "cuda".
+                automatically detect the available accelerator and use.
         """
         self._maybe_raise_error_if_group_offload_active(raise_error=True)
 
@@ -1118,6 +1181,11 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
 
         self.remove_all_hooks()
 
+        if device is None:
+            device = get_device()
+            if device == "cpu":
+                raise RuntimeError("`enable_model_cpu_offload` requires accelerator, but not found")
+
         torch_device = torch.device(device)
         device_index = torch_device.index
 
@@ -1135,9 +1203,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         self._offload_device = device
 
         self.to("cpu", silence_dtype_warnings=True)
-        device_mod = getattr(torch, device.type, None)
-        if hasattr(device_mod, "empty_cache") and device_mod.is_available():
-            device_mod.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+        empty_device_cache(device.type)
 
         all_model_components = {k: v for k, v in self.components.items() if isinstance(v, torch.nn.Module)}
 
@@ -1196,20 +1262,20 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         # make sure the model is in the same state as before calling it
         self.enable_model_cpu_offload(device=getattr(self, "_offload_device", "cuda"))
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models to CPU using 🤗 Accelerate, significantly reducing memory usage. When called, the state
         dicts of all `torch.nn.Module` components (except those in `self._exclude_from_cpu_offload`) are saved to CPU
-        and then moved to `torch.device('meta')` and loaded to GPU only when their specific submodule has its `forward`
-        method called. Offloading happens on a submodule basis. Memory savings are higher than with
+        and then moved to `torch.device('meta')` and loaded to accelerator only when their specific submodule has its
+        `forward` method called. Offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
 
         Arguments:
             gpu_id (`int`, *optional*):
                 The ID of the accelerator that shall be used in inference. If not specified, it will default to 0.
-            device (`torch.Device` or `str`, *optional*, defaults to "cuda"):
+            device (`torch.Device` or `str`, *optional*, defaults to None):
                 The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will
-                default to "cuda".
+                automatically detect the available accelerator and use.
         """
         self._maybe_raise_error_if_group_offload_active(raise_error=True)
 
@@ -1225,6 +1291,11 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_sequential_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_sequential_cpu_offload()`."
             )
 
+        if device is None:
+            device = get_device()
+            if device == "cpu":
+                raise RuntimeError("`enable_sequential_cpu_offload` requires accelerator, but not found")
+
         torch_device = torch.device(device)
         device_index = torch_device.index
 
@@ -1242,10 +1313,9 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         self._offload_device = device
 
         if self.device.type != "cpu":
+            orig_device_type = self.device.type
             self.to("cpu", silence_dtype_warnings=True)
-            device_mod = getattr(torch, self.device.type, None)
-            if hasattr(device_mod, "empty_cache") and device_mod.is_available():
-                device_mod.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+            empty_device_cache(orig_device_type)
 
         for name, model in self.components.items():
             if not isinstance(model, torch.nn.Module):
@@ -1628,6 +1698,8 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 signature_types[k] = (v.annotation,)
             elif get_origin(v.annotation) == Union:
                 signature_types[k] = get_args(v.annotation)
+            elif get_origin(v.annotation) in [List, Dict, list, dict]:
+                signature_types[k] = (v.annotation,)
             else:
                 logger.warning(f"cannot get type annotation for Parameter {k} of {cls}.")
         return signature_types
@@ -1990,7 +2062,7 @@ class StableDiffusionMixin:
         self.vae.disable_tiling()
 
     def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
-        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+        r"""Enables the FreeU mechanism as in https://huggingface.co/papers/2309.11497.
 
         The suffixes after the scaling factors represent the stages where they are being applied.