diffusers 0.32.1__py3-none-any.whl → 0.33.0__py3-none-any.whl

diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.32.1"
+__version__ = "0.33.0"
 
 from typing import TYPE_CHECKING
 
@@ -6,14 +6,20 @@ from .utils import (
     DIFFUSERS_SLOW_IMPORT,
     OptionalDependencyNotAvailable,
     _LazyModule,
+    is_accelerate_available,
+    is_bitsandbytes_available,
     is_flax_available,
+    is_gguf_available,
     is_k_diffusion_available,
     is_librosa_available,
     is_note_seq_available,
     is_onnx_available,
+    is_opencv_available,
+    is_optimum_quanto_available,
     is_scipy_available,
     is_sentencepiece_available,
     is_torch_available,
+    is_torchao_available,
     is_torchsde_available,
     is_transformers_available,
 )
@@ -28,10 +34,11 @@ from .utils import (
 
 _import_structure = {
     "configuration_utils": ["ConfigMixin"],
+    "hooks": [],
     "loaders": ["FromOriginalModelMixin"],
     "models": [],
     "pipelines": [],
-    "quantizers.quantization_config": ["BitsAndBytesConfig", "GGUFQuantizationConfig", "TorchAoConfig"],
+    "quantizers.quantization_config": [],
     "schedulers": [],
     "utils": [
         "OptionalDependencyNotAvailable",
@@ -53,6 +60,54 @@ _import_structure = {
     ],
 }
 
+try:
+    if not is_torch_available() and not is_accelerate_available() and not is_bitsandbytes_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_bitsandbytes_objects
+
+    _import_structure["utils.dummy_bitsandbytes_objects"] = [
+        name for name in dir(dummy_bitsandbytes_objects) if not name.startswith("_")
+    ]
+else:
+    _import_structure["quantizers.quantization_config"].append("BitsAndBytesConfig")
+
+try:
+    if not is_torch_available() and not is_accelerate_available() and not is_gguf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_gguf_objects
+
+    _import_structure["utils.dummy_gguf_objects"] = [
+        name for name in dir(dummy_gguf_objects) if not name.startswith("_")
+    ]
+else:
+    _import_structure["quantizers.quantization_config"].append("GGUFQuantizationConfig")
+
+try:
+    if not is_torch_available() and not is_accelerate_available() and not is_torchao_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torchao_objects
+
+    _import_structure["utils.dummy_torchao_objects"] = [
+        name for name in dir(dummy_torchao_objects) if not name.startswith("_")
+    ]
+else:
+    _import_structure["quantizers.quantization_config"].append("TorchAoConfig")
+
+try:
+    if not is_torch_available() and not is_accelerate_available() and not is_optimum_quanto_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_optimum_quanto_objects
+
+    _import_structure["utils.dummy_optimum_quanto_objects"] = [
+        name for name in dir(dummy_optimum_quanto_objects) if not name.startswith("_")
+    ]
+else:
+    _import_structure["quantizers.quantization_config"].append("QuantoConfig")
+
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()
@@ -75,6 +130,15 @@ except OptionalDependencyNotAvailable:
     _import_structure["utils.dummy_pt_objects"] = [name for name in dir(dummy_pt_objects) if not name.startswith("_")]
 
 else:
+    _import_structure["hooks"].extend(
+        [
+            "FasterCacheConfig",
+            "HookRegistry",
+            "PyramidAttentionBroadcastConfig",
+            "apply_faster_cache",
+            "apply_pyramid_attention_broadcast",
+        ]
+    )
     _import_structure["models"].extend(
         [
             "AllegroTransformer3DModel",
@@ -86,17 +150,24 @@ else:
             "AutoencoderKLCogVideoX",
             "AutoencoderKLHunyuanVideo",
             "AutoencoderKLLTXVideo",
+            "AutoencoderKLMagvit",
             "AutoencoderKLMochi",
             "AutoencoderKLTemporalDecoder",
+            "AutoencoderKLWan",
             "AutoencoderOobleck",
             "AutoencoderTiny",
+            "AutoModel",
+            "CacheMixin",
             "CogVideoXTransformer3DModel",
             "CogView3PlusTransformer2DModel",
+            "CogView4Transformer2DModel",
+            "ConsisIDTransformer3DModel",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
             "ControlNetUnionModel",
             "ControlNetXSAdapter",
             "DiTTransformer2DModel",
+            "EasyAnimateTransformer3DModel",
             "FluxControlNetModel",
             "FluxMultiControlNetModel",
             "FluxTransformer2DModel",
@@ -108,12 +179,14 @@ else:
             "Kandinsky3UNet",
             "LatteTransformer3DModel",
             "LTXVideoTransformer3DModel",
+            "Lumina2Transformer2DModel",
             "LuminaNextDiT2DModel",
             "MochiTransformer3DModel",
             "ModelMixin",
             "MotionAdapter",
             "MultiAdapter",
             "MultiControlNetModel",
+            "OmniGenTransformer2DModel",
             "PixArtTransformer2DModel",
             "PriorTransformer",
             "SanaTransformer2DModel",
@@ -126,6 +199,7 @@ else:
             "T2IAdapter",
             "T5FilmDecoder",
             "Transformer2DModel",
+            "TransformerTemporalModel",
             "UNet1DModel",
             "UNet2DConditionModel",
             "UNet2DModel",
@@ -135,6 +209,7 @@ else:
             "UNetSpatioTemporalConditionModel",
             "UVit2DModel",
             "VQModel",
+            "WanTransformer3DModel",
         ]
     )
     _import_structure["optimization"] = [
@@ -201,6 +276,7 @@ else:
             "RePaintScheduler",
             "SASolverScheduler",
             "SchedulerMixin",
+            "SCMScheduler",
             "ScoreSdeVeScheduler",
             "TCDScheduler",
             "UnCLIPScheduler",
@@ -275,7 +351,12 @@ else:
             "CogVideoXPipeline",
             "CogVideoXVideoToVideoPipeline",
             "CogView3PlusPipeline",
+            "CogView4ControlPipeline",
+            "CogView4Pipeline",
             "CycleDiffusionPipeline",
+            "EasyAnimateControlPipeline",
+            "EasyAnimateInpaintPipeline",
+            "EasyAnimatePipeline",
             "FluxControlImg2ImgPipeline",
             "FluxControlInpaintPipeline",
             "FluxControlNetImg2ImgPipeline",
@@ -290,6 +371,8 @@ else:
             "HunyuanDiTControlNetPipeline",
             "HunyuanDiTPAGPipeline",
             "HunyuanDiTPipeline",
+            "HunyuanSkyreelsImageToVideoPipeline",
+            "HunyuanVideoImageToVideoPipeline",
             "HunyuanVideoPipeline",
             "I2VGenXLPipeline",
             "IFImg2ImgPipeline",
@@ -324,13 +407,19 @@ else:
             "LDMTextToImagePipeline",
             "LEditsPPPipelineStableDiffusion",
             "LEditsPPPipelineStableDiffusionXL",
+            "LTXConditionPipeline",
             "LTXImageToVideoPipeline",
             "LTXPipeline",
+            "Lumina2Pipeline",
+            "Lumina2Text2ImgPipeline",
+            "LuminaPipeline",
             "LuminaText2ImgPipeline",
             "MarigoldDepthPipeline",
+            "MarigoldIntrinsicsPipeline",
             "MarigoldNormalsPipeline",
             "MochiPipeline",
             "MusicLDMPipeline",
+            "OmniGenPipeline",
             "PaintByExamplePipeline",
             "PIAPipeline",
             "PixArtAlphaPipeline",
@@ -339,6 +428,7 @@ else:
             "ReduxImageEncoder",
             "SanaPAGPipeline",
             "SanaPipeline",
+            "SanaSprintPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
             "ShapEPipeline",
@@ -419,12 +509,28 @@ else:
             "VersatileDiffusionTextToImagePipeline",
             "VideoToVideoSDPipeline",
             "VQDiffusionPipeline",
+            "WanImageToVideoPipeline",
+            "WanPipeline",
+            "WanVideoToVideoPipeline",
             "WuerstchenCombinedPipeline",
             "WuerstchenDecoderPipeline",
             "WuerstchenPriorPipeline",
         ]
     )
 
+try:
+    if not (is_torch_available() and is_transformers_available() and is_opencv_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_transformers_and_opencv_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_transformers_and_opencv_objects"] = [
+        name for name in dir(dummy_torch_and_transformers_and_opencv_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["pipelines"].extend(["ConsisIDPipeline"])
+
 try:
     if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
         raise OptionalDependencyNotAvailable()
@@ -570,7 +676,38 @@ else:
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     from .configuration_utils import ConfigMixin
-    from .quantizers.quantization_config import BitsAndBytesConfig, GGUFQuantizationConfig, TorchAoConfig
+
+    try:
+        if not is_bitsandbytes_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_bitsandbytes_objects import *
+    else:
+        from .quantizers.quantization_config import BitsAndBytesConfig
+
+    try:
+        if not is_gguf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_gguf_objects import *
+    else:
+        from .quantizers.quantization_config import GGUFQuantizationConfig
+
+    try:
+        if not is_torchao_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torchao_objects import *
+    else:
+        from .quantizers.quantization_config import TorchAoConfig
+
+    try:
+        if not is_optimum_quanto_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_optimum_quanto_objects import *
+    else:
+        from .quantizers.quantization_config import QuantoConfig
 
     try:
         if not is_onnx_available():
@@ -586,6 +723,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from .utils.dummy_pt_objects import *  # noqa F403
     else:
+        from .hooks import (
+            FasterCacheConfig,
+            HookRegistry,
+            PyramidAttentionBroadcastConfig,
+            apply_faster_cache,
+            apply_pyramid_attention_broadcast,
+        )
         from .models import (
             AllegroTransformer3DModel,
             AsymmetricAutoencoderKL,
@@ -596,17 +740,24 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLCogVideoX,
             AutoencoderKLHunyuanVideo,
             AutoencoderKLLTXVideo,
+            AutoencoderKLMagvit,
             AutoencoderKLMochi,
             AutoencoderKLTemporalDecoder,
+            AutoencoderKLWan,
             AutoencoderOobleck,
             AutoencoderTiny,
+            AutoModel,
+            CacheMixin,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,
+            CogView4Transformer2DModel,
+            ConsisIDTransformer3DModel,
             ConsistencyDecoderVAE,
             ControlNetModel,
             ControlNetUnionModel,
             ControlNetXSAdapter,
             DiTTransformer2DModel,
+            EasyAnimateTransformer3DModel,
             FluxControlNetModel,
             FluxMultiControlNetModel,
             FluxTransformer2DModel,
@@ -618,12 +769,14 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             Kandinsky3UNet,
             LatteTransformer3DModel,
             LTXVideoTransformer3DModel,
+            Lumina2Transformer2DModel,
             LuminaNextDiT2DModel,
             MochiTransformer3DModel,
             ModelMixin,
             MotionAdapter,
             MultiAdapter,
             MultiControlNetModel,
+            OmniGenTransformer2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
             SanaTransformer2DModel,
@@ -635,6 +788,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             T2IAdapter,
             T5FilmDecoder,
             Transformer2DModel,
+            TransformerTemporalModel,
             UNet1DModel,
             UNet2DConditionModel,
             UNet2DModel,
@@ -644,6 +798,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             UNetSpatioTemporalConditionModel,
             UVit2DModel,
             VQModel,
+            WanTransformer3DModel,
         )
         from .optimization import (
             get_constant_schedule,
@@ -709,6 +864,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             RePaintScheduler,
             SASolverScheduler,
             SchedulerMixin,
+            SCMScheduler,
             ScoreSdeVeScheduler,
             TCDScheduler,
             UnCLIPScheduler,
@@ -764,7 +920,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             CogVideoXPipeline,
             CogVideoXVideoToVideoPipeline,
             CogView3PlusPipeline,
+            CogView4ControlPipeline,
+            CogView4Pipeline,
             CycleDiffusionPipeline,
+            EasyAnimateControlPipeline,
+            EasyAnimateInpaintPipeline,
+            EasyAnimatePipeline,
             FluxControlImg2ImgPipeline,
             FluxControlInpaintPipeline,
             FluxControlNetImg2ImgPipeline,
@@ -779,6 +940,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             HunyuanDiTControlNetPipeline,
             HunyuanDiTPAGPipeline,
             HunyuanDiTPipeline,
+            HunyuanSkyreelsImageToVideoPipeline,
+            HunyuanVideoImageToVideoPipeline,
             HunyuanVideoPipeline,
             I2VGenXLPipeline,
             IFImg2ImgPipeline,
@@ -813,13 +976,19 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LDMTextToImagePipeline,
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
+            LTXConditionPipeline,
             LTXImageToVideoPipeline,
             LTXPipeline,
+            Lumina2Pipeline,
+            Lumina2Text2ImgPipeline,
+            LuminaPipeline,
             LuminaText2ImgPipeline,
             MarigoldDepthPipeline,
+            MarigoldIntrinsicsPipeline,
             MarigoldNormalsPipeline,
             MochiPipeline,
             MusicLDMPipeline,
+            OmniGenPipeline,
             PaintByExamplePipeline,
             PIAPipeline,
             PixArtAlphaPipeline,
@@ -828,6 +997,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ReduxImageEncoder,
             SanaPAGPipeline,
             SanaPipeline,
+            SanaSprintPipeline,
             SemanticStableDiffusionPipeline,
             ShapEImg2ImgPipeline,
             ShapEPipeline,
@@ -836,6 +1006,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableCascadeCombinedPipeline,
             StableCascadeDecoderPipeline,
             StableCascadePriorPipeline,
+            StableDiffusion3ControlNetInpaintingPipeline,
             StableDiffusion3ControlNetPipeline,
             StableDiffusion3Img2ImgPipeline,
             StableDiffusion3InpaintPipeline,
@@ -906,6 +1077,9 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VersatileDiffusionTextToImagePipeline,
             VideoToVideoSDPipeline,
             VQDiffusionPipeline,
+            WanImageToVideoPipeline,
+            WanPipeline,
+            WanVideoToVideoPipeline,
             WuerstchenCombinedPipeline,
             WuerstchenDecoderPipeline,
             WuerstchenPriorPipeline,
@@ -926,6 +1100,15 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .utils.dummy_torch_and_transformers_and_sentencepiece_objects import *  # noqa F403
     else:
         from .pipelines import KolorsImg2ImgPipeline, KolorsPAGPipeline, KolorsPipeline
+
+    try:
+        if not (is_torch_available() and is_transformers_available() and is_opencv_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_transformers_and_opencv_objects import *  # noqa F403
+    else:
+        from .pipelines import ConsisIDPipeline
+
     try:
         if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
             raise OptionalDependencyNotAvailable()

diffusers/configuration_utils.py

@@ -1,5 +1,5 @@
 # coding=utf-8
-# Copyright 2024 The HuggingFace Inc. team.
+# Copyright 2025 The HuggingFace Inc. team.
 # Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -24,10 +24,10 @@ import os
 import re
 from collections import OrderedDict
 from pathlib import Path
-from typing import Any, Dict, Tuple, Union
+from typing import Any, Dict, Optional, Tuple, Union
 
 import numpy as np
-from huggingface_hub import create_repo, hf_hub_download
+from huggingface_hub import DDUFEntry, create_repo, hf_hub_download
 from huggingface_hub.utils import (
     EntryNotFoundError,
     RepositoryNotFoundError,
@@ -35,6 +35,7 @@ from huggingface_hub.utils import (
     validate_hf_hub_args,
 )
 from requests import HTTPError
+from typing_extensions import Self
 
 from . import __version__
 from .utils import (
@@ -185,7 +186,9 @@ class ConfigMixin:
             )
 
     @classmethod
-    def from_config(cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs):
+    def from_config(
+        cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs
+    ) -> Union[Self, Tuple[Self, Dict[str, Any]]]:
         r"""
         Instantiate a Python class from a config dictionary.
 
@@ -347,6 +350,7 @@ class ConfigMixin:
         _ = kwargs.pop("mirror", None)
         subfolder = kwargs.pop("subfolder", None)
         user_agent = kwargs.pop("user_agent", {})
+        dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None)
 
         user_agent = {**user_agent, "file_type": "config"}
         user_agent = http_user_agent(user_agent)
@@ -358,8 +362,15 @@ class ConfigMixin:
                 "`self.config_name` is not defined. Note that one should not load a config from "
                 "`ConfigMixin`. Please make sure to define `config_name` in a class inheriting from `ConfigMixin`"
             )
-
-        if os.path.isfile(pretrained_model_name_or_path):
+        # Custom path for now
+        if dduf_entries:
+            if subfolder is not None:
+                raise ValueError(
+                    "DDUF file only allow for 1 level of directory (e.g transformer/model1/model.safetentors is not allowed). "
+                    "Please check the DDUF structure"
+                )
+            config_file = cls._get_config_file_from_dduf(pretrained_model_name_or_path, dduf_entries)
+        elif os.path.isfile(pretrained_model_name_or_path):
             config_file = pretrained_model_name_or_path
         elif os.path.isdir(pretrained_model_name_or_path):
             if subfolder is not None and os.path.isfile(
@@ -426,10 +437,8 @@ class ConfigMixin:
                     f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory "
                     f"containing a {cls.config_name} file"
                 )
-
         try:
-            # Load config dict
-            config_dict = cls._dict_from_json_file(config_file)
+            config_dict = cls._dict_from_json_file(config_file, dduf_entries=dduf_entries)
 
             commit_hash = extract_commit_hash(config_file)
         except (json.JSONDecodeError, UnicodeDecodeError):
@@ -552,9 +561,14 @@ class ConfigMixin:
         return init_dict, unused_kwargs, hidden_config_dict
 
     @classmethod
-    def _dict_from_json_file(cls, json_file: Union[str, os.PathLike]):
-        with open(json_file, "r", encoding="utf-8") as reader:
-            text = reader.read()
+    def _dict_from_json_file(
+        cls, json_file: Union[str, os.PathLike], dduf_entries: Optional[Dict[str, DDUFEntry]] = None
+    ):
+        if dduf_entries:
+            text = dduf_entries[json_file].read_text()
+        else:
+            with open(json_file, "r", encoding="utf-8") as reader:
+                text = reader.read()
         return json.loads(text)
 
     def __repr__(self):
@@ -616,6 +630,20 @@ class ConfigMixin:
         with open(json_file_path, "w", encoding="utf-8") as writer:
             writer.write(self.to_json_string())
 
+    @classmethod
+    def _get_config_file_from_dduf(cls, pretrained_model_name_or_path: str, dduf_entries: Dict[str, DDUFEntry]):
+        # paths inside a DDUF file must always be "/"
+        config_file = (
+            cls.config_name
+            if pretrained_model_name_or_path == ""
+            else "/".join([pretrained_model_name_or_path, cls.config_name])
+        )
+        if config_file not in dduf_entries:
+            raise ValueError(
+                f"We did not manage to find the file {config_file} in the dduf file. We only have the following files {dduf_entries.keys()}"
+            )
+        return config_file
+
 
 def register_to_config(init):
     r"""

diffusers/dependency_versions_table.py

@@ -9,7 +9,7 @@ deps = {
     "filelock": "filelock",
     "flax": "flax>=0.4.1",
     "hf-doc-builder": "hf-doc-builder>=0.3.0",
-    "huggingface-hub": "huggingface-hub>=0.23.2",
+    "huggingface-hub": "huggingface-hub>=0.27.0",
     "requests-mock": "requests-mock==1.10.0",
     "importlib_metadata": "importlib_metadata",
     "invisible-watermark": "invisible-watermark>=0.2.0",
@@ -29,18 +29,25 @@ deps = {
     "pytest-timeout": "pytest-timeout",
     "pytest-xdist": "pytest-xdist",
     "python": "python>=3.8.0",
-    "ruff": "ruff==0.1.5",
+    "ruff": "ruff==0.9.10",
     "safetensors": "safetensors>=0.3.1",
     "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",
     "GitPython": "GitPython<3.1.19",
     "scipy": "scipy",
     "onnx": "onnx",
+    "optimum_quanto": "optimum_quanto>=0.2.6",
+    "gguf": "gguf>=0.10.0",
+    "torchao": "torchao>=0.7.0",
+    "bitsandbytes": "bitsandbytes>=0.43.3",
     "regex": "regex!=2019.12.17",
     "requests": "requests",
     "tensorboard": "tensorboard",
+    "tiktoken": "tiktoken>=0.7.0",
     "torch": "torch>=1.4",
     "torchvision": "torchvision",
     "transformers": "transformers>=4.41.2",
     "urllib3": "urllib3<=2.0.0",
     "black": "black",
+    "phonemizer": "phonemizer",
+    "opencv-python": "opencv-python",
 }

diffusers/hooks/__init__.py

@@ -0,0 +1,9 @@
+from ..utils import is_torch_available
+
+
+if is_torch_available():
+    from .faster_cache import FasterCacheConfig, apply_faster_cache
+    from .group_offloading import apply_group_offloading
+    from .hooks import HookRegistry, ModelHook
+    from .layerwise_casting import apply_layerwise_casting, apply_layerwise_casting_hook
+    from .pyramid_attention_broadcast import PyramidAttentionBroadcastConfig, apply_pyramid_attention_broadcast