[Core] better support offloading when side loading is enabled.

src/diffusers/loaders.py

@@ -45,6 +45,7 @@
 
 if is_accelerate_available():
     from accelerate import init_empty_weights
+    from accelerate.hooks import CpuOffload, remove_hook_from_module
     from accelerate.utils import set_module_tensor_to_device
 
 logger = logging.get_logger(__name__)
@@ -763,6 +764,16 @@ def load_textual_inversion(
                 f" `{self.load_textual_inversion.__name__}`"
             )
 
+        # Remove any existing hooks.
+        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)
+                    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)
+
         cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
@@ -916,6 +927,12 @@ def load_textual_inversion(
         for token_id, embedding in token_ids_and_embeddings:
             self.text_encoder.get_input_embeddings().weight.data[token_id] = embedding
 
+        # offload back
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        else:
+            self.enable_sequential_cpu_offload()
+
 
 class LoraLoaderMixin:
     r"""
@@ -946,6 +963,16 @@ def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Di
             kwargs (`dict`, *optional*):
                 See [`~loaders.LoraLoaderMixin.lora_state_dict`].
         """
+        # Remove any existing hooks.
+        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)
+                    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)
+
         state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
         self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
         self.load_lora_into_text_encoder(
@@ -955,6 +982,12 @@ def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Di
             lora_scale=self.lora_scale,
         )
 
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        else:
+            self.enable_sequential_cpu_offload()
+
     @classmethod
     def lora_state_dict(
         cls,