diffusers 0.34.0__py3-none-any.whl → 0.35.1__py3-none-any.whl

diffusers/loaders/peft.py

@@ -61,6 +61,7 @@ _SET_ADAPTER_SCALE_FN_MAPPING = {
     "HunyuanVideoFramepackTransformer3DModel": lambda model_cls, weights: weights,
     "WanVACETransformer3DModel": lambda model_cls, weights: weights,
     "ChromaTransformer2DModel": lambda model_cls, weights: weights,
+    "QwenImageTransformer2DModel": lambda model_cls, weights: weights,
 }
 
 
@@ -163,6 +164,8 @@ class PeftAdapterMixin:
         from peft import inject_adapter_in_model, set_peft_model_state_dict
         from peft.tuners.tuners_utils import BaseTunerLayer
 
+        from ..hooks.group_offloading import _maybe_remove_and_reapply_group_offloading
+
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
@@ -243,20 +246,29 @@ class PeftAdapterMixin:
                     k.removeprefix(f"{prefix}."): v for k, v in network_alphas.items() if k in alpha_keys
                 }
 
-            # create LoraConfig
-            lora_config = _create_lora_config(state_dict, network_alphas, metadata, rank)
-
             # adapter_name
             if adapter_name is None:
                 adapter_name = get_adapter_name(self)
 
+            # create LoraConfig
+            lora_config = _create_lora_config(
+                state_dict,
+                network_alphas,
+                metadata,
+                rank,
+                model_state_dict=self.state_dict(),
+                adapter_name=adapter_name,
+            )
+
             # <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 `_pipeline`.
 
             # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
             # otherwise loading LoRA weights will lead to an error.
-            is_model_cpu_offload, is_sequential_cpu_offload = self._optionally_disable_offloading(_pipeline)
+            is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload = self._optionally_disable_offloading(
+                _pipeline
+            )
             peft_kwargs = {}
             if is_peft_version(">=", "0.13.1"):
                 peft_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
@@ -308,7 +320,9 @@ class PeftAdapterMixin:
                     # it to None
                     incompatible_keys = None
                 else:
-                    inject_adapter_in_model(lora_config, self, adapter_name=adapter_name, **peft_kwargs)
+                    inject_adapter_in_model(
+                        lora_config, self, adapter_name=adapter_name, state_dict=state_dict, **peft_kwargs
+                    )
                     incompatible_keys = set_peft_model_state_dict(self, state_dict, adapter_name, **peft_kwargs)
 
                     if self._prepare_lora_hotswap_kwargs is not None:
@@ -347,6 +361,10 @@ class PeftAdapterMixin:
                 _pipeline.enable_model_cpu_offload()
             elif is_sequential_cpu_offload:
                 _pipeline.enable_sequential_cpu_offload()
+            elif is_group_offload:
+                for component in _pipeline.components.values():
+                    if isinstance(component, torch.nn.Module):
+                        _maybe_remove_and_reapply_group_offloading(component)
             # Unsafe code />
 
         if prefix is not None and not state_dict:
@@ -681,11 +699,16 @@ class PeftAdapterMixin:
         if not USE_PEFT_BACKEND:
             raise ValueError("PEFT backend is required for `unload_lora()`.")
 
+        from ..hooks.group_offloading import _maybe_remove_and_reapply_group_offloading
         from ..utils import recurse_remove_peft_layers
 
         recurse_remove_peft_layers(self)
         if hasattr(self, "peft_config"):
             del self.peft_config
+        if hasattr(self, "_hf_peft_config_loaded"):
+            self._hf_peft_config_loaded = None
+
+        _maybe_remove_and_reapply_group_offloading(self)
 
     def disable_lora(self):
         """

diffusers/loaders/single_file_model.py

@@ -23,7 +23,8 @@ from typing_extensions import Self
 
 from .. import __version__
 from ..quantizers import DiffusersAutoQuantizer
-from ..utils import deprecate, is_accelerate_available, logging
+from ..utils import deprecate, is_accelerate_available, is_torch_version, logging
+from ..utils.torch_utils import empty_device_cache
 from .single_file_utils import (
     SingleFileComponentError,
     convert_animatediff_checkpoint_to_diffusers,
@@ -31,6 +32,7 @@ from .single_file_utils import (
     convert_autoencoder_dc_checkpoint_to_diffusers,
     convert_chroma_transformer_checkpoint_to_diffusers,
     convert_controlnet_checkpoint,
+    convert_cosmos_transformer_checkpoint_to_diffusers,
     convert_flux_transformer_checkpoint_to_diffusers,
     convert_hidream_transformer_to_diffusers,
     convert_hunyuan_video_transformer_to_diffusers,
@@ -60,8 +62,12 @@ logger = logging.get_logger(__name__)
 if is_accelerate_available():
     from accelerate import dispatch_model, init_empty_weights
 
-    from ..models.modeling_utils import load_model_dict_into_meta
+    from ..models.model_loading_utils import load_model_dict_into_meta
 
+if is_torch_version(">=", "1.9.0") and is_accelerate_available():
+    _LOW_CPU_MEM_USAGE_DEFAULT = True
+else:
+    _LOW_CPU_MEM_USAGE_DEFAULT = False
 
 SINGLE_FILE_LOADABLE_CLASSES = {
     "StableCascadeUNet": {
@@ -135,6 +141,10 @@ SINGLE_FILE_LOADABLE_CLASSES = {
         "checkpoint_mapping_fn": convert_wan_transformer_to_diffusers,
         "default_subfolder": "transformer",
     },
+    "WanVACETransformer3DModel": {
+        "checkpoint_mapping_fn": convert_wan_transformer_to_diffusers,
+        "default_subfolder": "transformer",
+    },
     "AutoencoderKLWan": {
         "checkpoint_mapping_fn": convert_wan_vae_to_diffusers,
         "default_subfolder": "vae",
@@ -143,9 +153,21 @@ SINGLE_FILE_LOADABLE_CLASSES = {
         "checkpoint_mapping_fn": convert_hidream_transformer_to_diffusers,
         "default_subfolder": "transformer",
     },
+    "CosmosTransformer3DModel": {
+        "checkpoint_mapping_fn": convert_cosmos_transformer_checkpoint_to_diffusers,
+        "default_subfolder": "transformer",
+    },
+    "QwenImageTransformer2DModel": {
+        "checkpoint_mapping_fn": lambda x: x,
+        "default_subfolder": "transformer",
+    },
 }
 
 
+def _should_convert_state_dict_to_diffusers(model_state_dict, checkpoint_state_dict):
+    return not set(model_state_dict.keys()).issubset(set(checkpoint_state_dict.keys()))
+
+
 def _get_single_file_loadable_mapping_class(cls):
     diffusers_module = importlib.import_module(__name__.split(".")[0])
     for loadable_class_str in SINGLE_FILE_LOADABLE_CLASSES:
@@ -218,6 +240,11 @@ class FromOriginalModelMixin:
             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.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 and
+                is_accelerate_available() 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.
             disable_mmap ('bool', *optional*, defaults to 'False'):
                 Whether to disable mmap when loading a Safetensors model. This option can perform better when the model
                 is on a network mount or hard drive, which may not handle the seeky-ness of mmap very well.
@@ -267,6 +294,7 @@ class FromOriginalModelMixin:
         config_revision = kwargs.pop("config_revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
         quantization_config = kwargs.pop("quantization_config", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
         device = kwargs.pop("device", None)
         disable_mmap = kwargs.pop("disable_mmap", False)
 
@@ -371,19 +399,23 @@ class FromOriginalModelMixin:
             model_kwargs = {k: kwargs.get(k) for k in kwargs if k in expected_kwargs or k in optional_kwargs}
             diffusers_model_config.update(model_kwargs)
 
+        ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+        with ctx():
+            model = cls.from_config(diffusers_model_config)
+
         checkpoint_mapping_kwargs = _get_mapping_function_kwargs(checkpoint_mapping_fn, **kwargs)
-        diffusers_format_checkpoint = checkpoint_mapping_fn(
-            config=diffusers_model_config, checkpoint=checkpoint, **checkpoint_mapping_kwargs
-        )
+
+        if _should_convert_state_dict_to_diffusers(model.state_dict(), checkpoint):
+            diffusers_format_checkpoint = checkpoint_mapping_fn(
+                config=diffusers_model_config, checkpoint=checkpoint, **checkpoint_mapping_kwargs
+            )
+        else:
+            diffusers_format_checkpoint = checkpoint
+
         if not diffusers_format_checkpoint:
             raise SingleFileComponentError(
                 f"Failed to load {mapping_class_name}. Weights for this component appear to be missing in the checkpoint."
             )
-
-        ctx = init_empty_weights if is_accelerate_available() else nullcontext
-        with ctx():
-            model = cls.from_config(diffusers_model_config)
-
         # Check if `_keep_in_fp32_modules` is not None
         use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (
             (torch_dtype == torch.float16) or hasattr(hf_quantizer, "use_keep_in_fp32_modules")
@@ -405,7 +437,7 @@ class FromOriginalModelMixin:
             )
 
         device_map = None
-        if is_accelerate_available():
+        if low_cpu_mem_usage:
             param_device = torch.device(device) if device else torch.device("cpu")
             empty_state_dict = model.state_dict()
             unexpected_keys = [
@@ -421,6 +453,7 @@ class FromOriginalModelMixin:
                 keep_in_fp32_modules=keep_in_fp32_modules,
                 unexpected_keys=unexpected_keys,
             )
+            empty_device_cache()
         else:
             _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)
 

diffusers/loaders/single_file_utils.py

@@ -46,6 +46,7 @@ from ..utils import (
 )
 from ..utils.constants import DIFFUSERS_REQUEST_TIMEOUT
 from ..utils.hub_utils import _get_model_file
+from ..utils.torch_utils import empty_device_cache
 
 
 if is_transformers_available():
@@ -54,11 +55,12 @@ if is_transformers_available():
 if is_accelerate_available():
     from accelerate import init_empty_weights
 
-    from ..models.modeling_utils import load_model_dict_into_meta
+    from ..models.model_loading_utils import load_model_dict_into_meta
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 CHECKPOINT_KEY_NAMES = {
+    "v1": "model.diffusion_model.output_blocks.11.0.skip_connection.weight",
     "v2": "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight",
     "xl_base": "conditioner.embedders.1.model.transformer.resblocks.9.mlp.c_proj.bias",
     "xl_refiner": "conditioner.embedders.0.model.transformer.resblocks.9.mlp.c_proj.bias",
@@ -126,7 +128,18 @@ CHECKPOINT_KEY_NAMES = {
     ],
     "wan": ["model.diffusion_model.head.modulation", "head.modulation"],
     "wan_vae": "decoder.middle.0.residual.0.gamma",
+    "wan_vace": "vace_blocks.0.after_proj.bias",
     "hidream": "double_stream_blocks.0.block.adaLN_modulation.1.bias",
+    "cosmos-1.0": [
+        "net.x_embedder.proj.1.weight",
+        "net.blocks.block1.blocks.0.block.attn.to_q.0.weight",
+        "net.extra_pos_embedder.pos_emb_h",
+    ],
+    "cosmos-2.0": [
+        "net.x_embedder.proj.1.weight",
+        "net.blocks.0.self_attn.q_proj.weight",
+        "net.pos_embedder.dim_spatial_range",
+    ],
 }
 
 DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
@@ -192,7 +205,17 @@ DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
     "wan-t2v-1.3B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"},
     "wan-t2v-14B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-T2V-14B-Diffusers"},
     "wan-i2v-14B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"},
+    "wan-vace-1.3B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-VACE-1.3B-diffusers"},
+    "wan-vace-14B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-VACE-14B-diffusers"},
     "hidream": {"pretrained_model_name_or_path": "HiDream-ai/HiDream-I1-Dev"},
+    "cosmos-1.0-t2w-7B": {"pretrained_model_name_or_path": "nvidia/Cosmos-1.0-Diffusion-7B-Text2World"},
+    "cosmos-1.0-t2w-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-1.0-Diffusion-14B-Text2World"},
+    "cosmos-1.0-v2w-7B": {"pretrained_model_name_or_path": "nvidia/Cosmos-1.0-Diffusion-7B-Video2World"},
+    "cosmos-1.0-v2w-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-1.0-Diffusion-14B-Video2World"},
+    "cosmos-2.0-t2i-2B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-2B-Text2Image"},
+    "cosmos-2.0-t2i-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-14B-Text2Image"},
+    "cosmos-2.0-v2w-2B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-2B-Video2World"},
+    "cosmos-2.0-v2w-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-14B-Video2World"},
 }
 
 # Use to configure model sample size when original config is provided
@@ -698,17 +721,44 @@ def infer_diffusers_model_type(checkpoint):
         else:
             target_key = "patch_embedding.weight"
 
-        if checkpoint[target_key].shape[0] == 1536:
+        if CHECKPOINT_KEY_NAMES["wan_vace"] in checkpoint:
+            if checkpoint[target_key].shape[0] == 1536:
+                model_type = "wan-vace-1.3B"
+            elif checkpoint[target_key].shape[0] == 5120:
+                model_type = "wan-vace-14B"
+
+        elif checkpoint[target_key].shape[0] == 1536:
             model_type = "wan-t2v-1.3B"
         elif checkpoint[target_key].shape[0] == 5120 and checkpoint[target_key].shape[1] == 16:
             model_type = "wan-t2v-14B"
         else:
             model_type = "wan-i2v-14B"
+
     elif CHECKPOINT_KEY_NAMES["wan_vae"] in checkpoint:
         # All Wan models use the same VAE so we can use the same default model repo to fetch the config
         model_type = "wan-t2v-14B"
+
     elif CHECKPOINT_KEY_NAMES["hidream"] in checkpoint:
         model_type = "hidream"
+
+    elif all(key in checkpoint for key in CHECKPOINT_KEY_NAMES["cosmos-1.0"]):
+        x_embedder_shape = checkpoint[CHECKPOINT_KEY_NAMES["cosmos-1.0"][0]].shape
+        if x_embedder_shape[1] == 68:
+            model_type = "cosmos-1.0-t2w-7B" if x_embedder_shape[0] == 4096 else "cosmos-1.0-t2w-14B"
+        elif x_embedder_shape[1] == 72:
+            model_type = "cosmos-1.0-v2w-7B" if x_embedder_shape[0] == 4096 else "cosmos-1.0-v2w-14B"
+        else:
+            raise ValueError(f"Unexpected x_embedder shape: {x_embedder_shape} when loading Cosmos 1.0 model.")
+
+    elif all(key in checkpoint for key in CHECKPOINT_KEY_NAMES["cosmos-2.0"]):
+        x_embedder_shape = checkpoint[CHECKPOINT_KEY_NAMES["cosmos-2.0"][0]].shape
+        if x_embedder_shape[1] == 68:
+            model_type = "cosmos-2.0-t2i-2B" if x_embedder_shape[0] == 2048 else "cosmos-2.0-t2i-14B"
+        elif x_embedder_shape[1] == 72:
+            model_type = "cosmos-2.0-v2w-2B" if x_embedder_shape[0] == 2048 else "cosmos-2.0-v2w-14B"
+        else:
+            raise ValueError(f"Unexpected x_embedder shape: {x_embedder_shape} when loading Cosmos 2.0 model.")
+
     else:
         model_type = "v1"
 
@@ -1641,6 +1691,7 @@ def create_diffusers_clip_model_from_ldm(
 
     if is_accelerate_available():
         load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
+        empty_device_cache()
     else:
         model.load_state_dict(diffusers_format_checkpoint, strict=False)
 
@@ -2100,6 +2151,7 @@ def create_diffusers_t5_model_from_checkpoint(
 
     if is_accelerate_available():
         load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
+        empty_device_cache()
     else:
         model.load_state_dict(diffusers_format_checkpoint)
 
@@ -3093,6 +3145,9 @@ def convert_wan_transformer_to_diffusers(checkpoint, **kwargs):
         "img_emb.proj.1": "condition_embedder.image_embedder.ff.net.0.proj",
         "img_emb.proj.3": "condition_embedder.image_embedder.ff.net.2",
         "img_emb.proj.4": "condition_embedder.image_embedder.norm2",
+        # For the VACE model
+        "before_proj": "proj_in",
+        "after_proj": "proj_out",
     }
 
     for key in list(checkpoint.keys()):
@@ -3479,3 +3534,116 @@ def convert_chroma_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
     converted_state_dict["proj_out.bias"] = checkpoint.pop("final_layer.linear.bias")
 
     return converted_state_dict
+
+
+def convert_cosmos_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
+    converted_state_dict = {key: checkpoint.pop(key) for key in list(checkpoint.keys())}
+
+    def remove_keys_(key: str, state_dict):
+        state_dict.pop(key)
+
+    def rename_transformer_blocks_(key: str, state_dict):
+        block_index = int(key.split(".")[1].removeprefix("block"))
+        new_key = key
+        old_prefix = f"blocks.block{block_index}"
+        new_prefix = f"transformer_blocks.{block_index}"
+        new_key = new_prefix + new_key.removeprefix(old_prefix)
+        state_dict[new_key] = state_dict.pop(key)
+
+    TRANSFORMER_KEYS_RENAME_DICT_COSMOS_1_0 = {
+        "t_embedder.1": "time_embed.t_embedder",
+        "affline_norm": "time_embed.norm",
+        ".blocks.0.block.attn": ".attn1",
+        ".blocks.1.block.attn": ".attn2",
+        ".blocks.2.block": ".ff",
+        ".blocks.0.adaLN_modulation.1": ".norm1.linear_1",
+        ".blocks.0.adaLN_modulation.2": ".norm1.linear_2",
+        ".blocks.1.adaLN_modulation.1": ".norm2.linear_1",
+        ".blocks.1.adaLN_modulation.2": ".norm2.linear_2",
+        ".blocks.2.adaLN_modulation.1": ".norm3.linear_1",
+        ".blocks.2.adaLN_modulation.2": ".norm3.linear_2",
+        "to_q.0": "to_q",
+        "to_q.1": "norm_q",
+        "to_k.0": "to_k",
+        "to_k.1": "norm_k",
+        "to_v.0": "to_v",
+        "layer1": "net.0.proj",
+        "layer2": "net.2",
+        "proj.1": "proj",
+        "x_embedder": "patch_embed",
+        "extra_pos_embedder": "learnable_pos_embed",
+        "final_layer.adaLN_modulation.1": "norm_out.linear_1",
+        "final_layer.adaLN_modulation.2": "norm_out.linear_2",
+        "final_layer.linear": "proj_out",
+    }
+
+    TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_1_0 = {
+        "blocks.block": rename_transformer_blocks_,
+        "logvar.0.freqs": remove_keys_,
+        "logvar.0.phases": remove_keys_,
+        "logvar.1.weight": remove_keys_,
+        "pos_embedder.seq": remove_keys_,
+    }
+
+    TRANSFORMER_KEYS_RENAME_DICT_COSMOS_2_0 = {
+        "t_embedder.1": "time_embed.t_embedder",
+        "t_embedding_norm": "time_embed.norm",
+        "blocks": "transformer_blocks",
+        "adaln_modulation_self_attn.1": "norm1.linear_1",
+        "adaln_modulation_self_attn.2": "norm1.linear_2",
+        "adaln_modulation_cross_attn.1": "norm2.linear_1",
+        "adaln_modulation_cross_attn.2": "norm2.linear_2",
+        "adaln_modulation_mlp.1": "norm3.linear_1",
+        "adaln_modulation_mlp.2": "norm3.linear_2",
+        "self_attn": "attn1",
+        "cross_attn": "attn2",
+        "q_proj": "to_q",
+        "k_proj": "to_k",
+        "v_proj": "to_v",
+        "output_proj": "to_out.0",
+        "q_norm": "norm_q",
+        "k_norm": "norm_k",
+        "mlp.layer1": "ff.net.0.proj",
+        "mlp.layer2": "ff.net.2",
+        "x_embedder.proj.1": "patch_embed.proj",
+        "final_layer.adaln_modulation.1": "norm_out.linear_1",
+        "final_layer.adaln_modulation.2": "norm_out.linear_2",
+        "final_layer.linear": "proj_out",
+    }
+
+    TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_2_0 = {
+        "accum_video_sample_counter": remove_keys_,
+        "accum_image_sample_counter": remove_keys_,
+        "accum_iteration": remove_keys_,
+        "accum_train_in_hours": remove_keys_,
+        "pos_embedder.seq": remove_keys_,
+        "pos_embedder.dim_spatial_range": remove_keys_,
+        "pos_embedder.dim_temporal_range": remove_keys_,
+        "_extra_state": remove_keys_,
+    }
+
+    PREFIX_KEY = "net."
+    if "net.blocks.block1.blocks.0.block.attn.to_q.0.weight" in checkpoint:
+        TRANSFORMER_KEYS_RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT_COSMOS_1_0
+        TRANSFORMER_SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_1_0
+    else:
+        TRANSFORMER_KEYS_RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT_COSMOS_2_0
+        TRANSFORMER_SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP_COSMOS_2_0
+
+    state_dict_keys = list(converted_state_dict.keys())
+    for key in state_dict_keys:
+        new_key = key[:]
+        if new_key.startswith(PREFIX_KEY):
+            new_key = new_key.removeprefix(PREFIX_KEY)
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        converted_state_dict[new_key] = converted_state_dict.pop(key)
+
+    state_dict_keys = list(converted_state_dict.keys())
+    for key in state_dict_keys:
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, converted_state_dict)
+
+    return converted_state_dict

diffusers/loaders/transformer_flux.py

@@ -17,12 +17,10 @@ from ..models.embeddings import (
     ImageProjection,
     MultiIPAdapterImageProjection,
 )
-from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
-from ..utils import (
-    is_accelerate_available,
-    is_torch_version,
-    logging,
-)
+from ..models.model_loading_utils import load_model_dict_into_meta
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
+from ..utils import is_accelerate_available, is_torch_version, logging
+from ..utils.torch_utils import empty_device_cache
 
 
 if is_accelerate_available():
@@ -84,13 +82,12 @@ class FluxTransformer2DLoadersMixin:
         else:
             device_map = {"": self.device}
             load_model_dict_into_meta(image_projection, updated_state_dict, device_map=device_map, dtype=self.dtype)
+            empty_device_cache()
 
         return image_projection
 
     def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=_LOW_CPU_MEM_USAGE_DEFAULT):
-        from ..models.attention_processor import (
-            FluxIPAdapterJointAttnProcessor2_0,
-        )
+        from ..models.transformers.transformer_flux import FluxIPAdapterAttnProcessor
 
         if low_cpu_mem_usage:
             if is_accelerate_available():
@@ -122,7 +119,7 @@ class FluxTransformer2DLoadersMixin:
             else:
                 cross_attention_dim = self.config.joint_attention_dim
                 hidden_size = self.inner_dim
-                attn_processor_class = FluxIPAdapterJointAttnProcessor2_0
+                attn_processor_class = FluxIPAdapterAttnProcessor
                 num_image_text_embeds = []
                 for state_dict in state_dicts:
                     if "proj.weight" in state_dict["image_proj"]:
@@ -158,6 +155,8 @@ class FluxTransformer2DLoadersMixin:
 
                 key_id += 1
 
+        empty_device_cache()
+
         return attn_procs
 
     def _load_ip_adapter_weights(self, state_dicts, low_cpu_mem_usage=_LOW_CPU_MEM_USAGE_DEFAULT):

diffusers/loaders/transformer_sd3.py

@@ -16,8 +16,10 @@ from typing import Dict
 
 from ..models.attention_processor import SD3IPAdapterJointAttnProcessor2_0
 from ..models.embeddings import IPAdapterTimeImageProjection
-from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+from ..models.model_loading_utils import load_model_dict_into_meta
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
 from ..utils import is_accelerate_available, is_torch_version, logging
+from ..utils.torch_utils import empty_device_cache
 
 
 logger = logging.get_logger(__name__)
@@ -80,6 +82,8 @@ class SD3Transformer2DLoadersMixin:
                     attn_procs[name], layer_state_dict[idx], device_map=device_map, dtype=self.dtype
                 )
 
+        empty_device_cache()
+
         return attn_procs
 
     def _convert_ip_adapter_image_proj_to_diffusers(
@@ -147,6 +151,7 @@ class SD3Transformer2DLoadersMixin:
         else:
             device_map = {"": self.device}
             load_model_dict_into_meta(image_proj, updated_state_dict, device_map=device_map, dtype=self.dtype)
+            empty_device_cache()
 
         return image_proj
 

diffusers/loaders/unet.py

@@ -30,7 +30,8 @@ from ..models.embeddings import (
     IPAdapterPlusImageProjection,
     MultiIPAdapterImageProjection,
 )
-from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta, load_state_dict
+from ..models.model_loading_utils import load_model_dict_into_meta
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_state_dict
 from ..utils import (
     USE_PEFT_BACKEND,
     _get_model_file,
@@ -43,6 +44,7 @@ from ..utils import (
     is_torch_version,
     logging,
 )
+from ..utils.torch_utils import empty_device_cache
 from .lora_base import _func_optionally_disable_offloading
 from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE, TEXT_ENCODER_NAME, UNET_NAME
 from .utils import AttnProcsLayers
@@ -131,6 +133,8 @@ class UNet2DConditionLoadersMixin:
         )
         ```
         """
+        from ..hooks.group_offloading import _maybe_remove_and_reapply_group_offloading
+
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
@@ -203,6 +207,7 @@ class UNet2DConditionLoadersMixin:
         is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys())
         is_model_cpu_offload = False
         is_sequential_cpu_offload = False
+        is_group_offload = False
 
         if is_lora:
             deprecation_message = "Using the `load_attn_procs()` method has been deprecated and will be removed in a future version. Please use `load_lora_adapter()`."
@@ -211,7 +216,7 @@ class UNet2DConditionLoadersMixin:
         if is_custom_diffusion:
             attn_processors = self._process_custom_diffusion(state_dict=state_dict)
         elif is_lora:
-            is_model_cpu_offload, is_sequential_cpu_offload = self._process_lora(
+            is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload = self._process_lora(
                 state_dict=state_dict,
                 unet_identifier_key=self.unet_name,
                 network_alphas=network_alphas,
@@ -230,7 +235,9 @@ class UNet2DConditionLoadersMixin:
 
         # For LoRA, the UNet is already offloaded at this stage as it is handled inside `_process_lora`.
         if is_custom_diffusion and _pipeline is not None:
-            is_model_cpu_offload, is_sequential_cpu_offload = self._optionally_disable_offloading(_pipeline=_pipeline)
+            is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload = self._optionally_disable_offloading(
+                _pipeline=_pipeline
+            )
 
             # only custom diffusion needs to set attn processors
             self.set_attn_processor(attn_processors)
@@ -241,6 +248,10 @@ class UNet2DConditionLoadersMixin:
             _pipeline.enable_model_cpu_offload()
         elif is_sequential_cpu_offload:
             _pipeline.enable_sequential_cpu_offload()
+        elif is_group_offload:
+            for component in _pipeline.components.values():
+                if isinstance(component, torch.nn.Module):
+                    _maybe_remove_and_reapply_group_offloading(component)
         # Unsafe code />
 
     def _process_custom_diffusion(self, state_dict):
@@ -307,6 +318,7 @@ class UNet2DConditionLoadersMixin:
 
         is_model_cpu_offload = False
         is_sequential_cpu_offload = False
+        is_group_offload = False
         state_dict_to_be_used = unet_state_dict if len(unet_state_dict) > 0 else state_dict
 
         if len(state_dict_to_be_used) > 0:
@@ -356,7 +368,9 @@ class UNet2DConditionLoadersMixin:
 
             # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
             # otherwise loading LoRA weights will lead to an error
-            is_model_cpu_offload, is_sequential_cpu_offload = self._optionally_disable_offloading(_pipeline)
+            is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload = self._optionally_disable_offloading(
+                _pipeline
+            )
             peft_kwargs = {}
             if is_peft_version(">=", "0.13.1"):
                 peft_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
@@ -389,7 +403,7 @@ class UNet2DConditionLoadersMixin:
             if warn_msg:
                 logger.warning(warn_msg)
 
-        return is_model_cpu_offload, is_sequential_cpu_offload
+        return is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload
 
     @classmethod
     # Copied from diffusers.loaders.lora_base.LoraBaseMixin._optionally_disable_offloading
@@ -741,6 +755,7 @@ class UNet2DConditionLoadersMixin:
         else:
             device_map = {"": self.device}
             load_model_dict_into_meta(image_projection, updated_state_dict, device_map=device_map, dtype=self.dtype)
+            empty_device_cache()
 
         return image_projection
 
@@ -838,6 +853,8 @@ class UNet2DConditionLoadersMixin:
 
                 key_id += 2
 
+        empty_device_cache()
+
         return attn_procs
 
     def _load_ip_adapter_weights(self, state_dicts, low_cpu_mem_usage=_LOW_CPU_MEM_USAGE_DEFAULT):