Refactor execution device & cpu offload (#4114)

* create general cpu offload & execution device

* Remove boiler plate

* finish

* kp

* Correct offload more pipelines

* up

* Update src/diffusers/pipelines/pipeline_utils.py

* make style

* up

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py

@@ -110,27 +110,6 @@ class KandinskyV22Pipeline(DiffusionPipeline):
         latents = latents * scheduler.init_noise_sigma
         return latents
 
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [
-            self.unet,
-            self.movq,
-        ]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
@@ -156,25 +135,6 @@ class KandinskyV22Pipeline(DiffusionPipeline):
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet.py

@@ -150,28 +150,6 @@ class KandinskyV22ControlnetPipeline(DiffusionPipeline):
         latents = latents * scheduler.init_noise_sigma
         return latents
 
-    # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.KandinskyV22Pipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [
-            self.unet,
-            self.movq,
-        ]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
     # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.KandinskyV22Pipeline.enable_model_cpu_offload
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
@@ -198,25 +176,6 @@ class KandinskyV22ControlnetPipeline(DiffusionPipeline):
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet_img2img.py

@@ -205,28 +205,6 @@ class KandinskyV22ControlnetImg2ImgPipeline(DiffusionPipeline):
 
         return latents
 
-    # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.KandinskyV22Pipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [
-            self.unet,
-            self.movq,
-        ]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
     # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.KandinskyV22Pipeline.enable_model_cpu_offload
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
@@ -253,25 +231,6 @@ class KandinskyV22ControlnetImg2ImgPipeline(DiffusionPipeline):
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py

@@ -178,28 +178,6 @@ class KandinskyV22Img2ImgPipeline(DiffusionPipeline):
 
         return latents
 
-    # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.KandinskyV22Pipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [
-            self.unet,
-            self.movq,
-        ]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
     # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.KandinskyV22Pipeline.enable_model_cpu_offload
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
@@ -226,25 +204,6 @@ class KandinskyV22Img2ImgPipeline(DiffusionPipeline):
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpainting.py

@@ -277,28 +277,6 @@ class KandinskyV22InpaintPipeline(DiffusionPipeline):
         latents = latents * scheduler.init_noise_sigma
         return latents
 
-    # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.KandinskyV22Pipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [
-            self.unet,
-            self.movq,
-        ]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
     # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.KandinskyV22Pipeline.enable_model_cpu_offload
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
@@ -325,25 +303,6 @@ class KandinskyV22InpaintPipeline(DiffusionPipeline):
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py

@@ -8,7 +8,6 @@ from ...models import PriorTransformer
 from ...pipelines import DiffusionPipeline
 from ...schedulers import UnCLIPScheduler
 from ...utils import (
-    is_accelerate_available,
     logging,
     randn_tensor,
     replace_example_docstring,
@@ -105,6 +104,8 @@ class KandinskyV22PriorPipeline(DiffusionPipeline):
             A image_processor to be used to preprocess image from clip.
     """
 
+    _exclude_from_cpu_offload = ["prior"]
+
     def __init__(
         self,
         prior: PriorTransformer,
@@ -254,47 +255,6 @@ class KandinskyV22PriorPipeline(DiffusionPipeline):
         zero_image_emb = zero_image_emb.repeat(batch_size, 1)
         return zero_image_emb
 
-    # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [
-            self.image_encoder,
-            self.text_encoder,
-        ]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
-    @property
-    # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if self.device != torch.device("meta") or not hasattr(self.text_encoder, "_hf_hook"):
-            return self.device
-        for module in self.text_encoder.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline._encode_prompt
     def _encode_prompt(
         self,

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior_emb2emb.py

@@ -8,7 +8,6 @@ from ...models import PriorTransformer
 from ...pipelines import DiffusionPipeline
 from ...schedulers import UnCLIPScheduler
 from ...utils import (
-    is_accelerate_available,
     logging,
     randn_tensor,
     replace_example_docstring,
@@ -121,6 +120,8 @@ class KandinskyV22PriorEmb2EmbPipeline(DiffusionPipeline):
             A scheduler to be used in combination with `prior` to generate image embedding.
     """
 
+    _exclude_from_cpu_offload = ["prior"]
+
     def __init__(
         self,
         prior: PriorTransformer,
@@ -293,47 +294,6 @@ class KandinskyV22PriorEmb2EmbPipeline(DiffusionPipeline):
         zero_image_emb = zero_image_emb.repeat(batch_size, 1)
         return zero_image_emb
 
-    # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [
-            self.image_encoder,
-            self.text_encoder,
-        ]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
-    @property
-    # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if self.device != torch.device("meta") or not hasattr(self.text_encoder, "_hf_hook"):
-            return self.device
-        for module in self.text_encoder.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline._encode_prompt
     def _encode_prompt(
         self,

src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py

@@ -128,11 +128,11 @@ class LDMTextToImagePipeline(DiffusionPipeline):
             uncond_input = self.tokenizer(
                 [""] * batch_size, padding="max_length", max_length=77, truncation=True, return_tensors="pt"
             )
-            negative_prompt_embeds = self.bert(uncond_input.input_ids.to(self.device))[0]
+            negative_prompt_embeds = self.bert(uncond_input.input_ids.to(self._execution_device))[0]
 
         # get prompt text embeddings
         text_input = self.tokenizer(prompt, padding="max_length", max_length=77, truncation=True, return_tensors="pt")
-        prompt_embeds = self.bert(text_input.input_ids.to(self.device))[0]
+        prompt_embeds = self.bert(text_input.input_ids.to(self._execution_device))[0]
 
         # get the initial random noise unless the user supplied it
         latents_shape = (batch_size, self.unet.config.in_channels, height // 8, width // 8)
@@ -143,11 +143,13 @@ class LDMTextToImagePipeline(DiffusionPipeline):
             )
 
         if latents is None:
-            latents = randn_tensor(latents_shape, generator=generator, device=self.device, dtype=prompt_embeds.dtype)
+            latents = randn_tensor(
+                latents_shape, generator=generator, device=self._execution_device, dtype=prompt_embeds.dtype
+            )
         else:
             if latents.shape != latents_shape:
                 raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
-        latents = latents.to(self.device)
+        latents = latents.to(self._execution_device)
 
         self.scheduler.set_timesteps(num_inference_steps)
 

src/diffusers/pipelines/paint_by_example/pipeline_paint_by_example.py

@@ -21,8 +21,6 @@ import PIL
 import torch
 from transformers import CLIPImageProcessor
 
-from diffusers.utils import is_accelerate_available
-
 from ...image_processor import VaeImageProcessor
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
@@ -189,44 +187,6 @@ class PaintByExamplePipeline(DiffusionPipeline):
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
-        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
-        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        for cpu_offloaded_model in [self.unet, self.vae, self.image_encoder]:
-            cpu_offload(cpu_offloaded_model, execution_device=device)
-
-        if self.safety_checker is not None:
-            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
-
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:

src/diffusers/pipelines/pipeline_utils.py

@@ -474,6 +474,7 @@ class DiffusionPipeline(ConfigMixin):
     """
     config_name = "model_index.json"
     _optional_components = []
+    _exclude_from_cpu_offload = []
 
     def register_modules(self, **kwargs):
         # import it here to avoid circular import
@@ -1086,6 +1087,60 @@ class DiffusionPipeline(ConfigMixin):
     def name_or_path(self) -> str:
         return getattr(self.config, "_name_or_path", None)
 
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        [`~DiffusionPipeline.enable_sequential_cpu_offload`] the execution device can only be inferred from
+        Accelerate's module hooks.
+        """
+        for name, model in self.components.items():
+            if not isinstance(model, torch.nn.Module) or name in self._exclude_from_cpu_offload:
+                continue
+
+            if not hasattr(model, "_hf_hook"):
+                return self.device
+            for module in model.modules():
+                if (
+                    hasattr(module, "_hf_hook")
+                    and hasattr(module._hf_hook, "execution_device")
+                    and module._hf_hook.execution_device is not None
+                ):
+                    return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def enable_sequential_cpu_offload(self, gpu_id: int = 0, device: Union[torch.device, str] = "cuda"):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+        if device == "cuda":
+            device = torch.device(f"{device}:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        for name, model in self.components.items():
+            if not isinstance(model, torch.nn.Module):
+                continue
+
+            if name in self._exclude_from_cpu_offload:
+                model.to(device)
+            else:
+                # make sure to offload buffers if not all high level weights
+                # are of type nn.Module
+                offload_buffers = len(model._parameters) > 0
+                cpu_offload(model, device, offload_buffers=offload_buffers)
+
     @classmethod
     def download(cls, pretrained_model_name, **kwargs) -> Union[str, os.PathLike]:
         r"""

src/diffusers/pipelines/repaint/pipeline_repaint.py

@@ -132,9 +132,9 @@ class RePaintPipeline(DiffusionPipeline):
         original_image = image
 
         original_image = _preprocess_image(original_image)
-        original_image = original_image.to(device=self.device, dtype=self.unet.dtype)
+        original_image = original_image.to(device=self._execution_device, dtype=self.unet.dtype)
         mask_image = _preprocess_mask(mask_image)
-        mask_image = mask_image.to(device=self.device, dtype=self.unet.dtype)
+        mask_image = mask_image.to(device=self._execution_device, dtype=self.unet.dtype)
 
         batch_size = original_image.shape[0]
 
@@ -146,10 +146,10 @@ class RePaintPipeline(DiffusionPipeline):
             )
 
         image_shape = original_image.shape
-        image = randn_tensor(image_shape, generator=generator, device=self.device, dtype=self.unet.dtype)
+        image = randn_tensor(image_shape, generator=generator, device=self._execution_device, dtype=self.unet.dtype)
 
         # set step values
-        self.scheduler.set_timesteps(num_inference_steps, jump_length, jump_n_sample, self.device)
+        self.scheduler.set_timesteps(num_inference_steps, jump_length, jump_n_sample, self._execution_device)
         self.scheduler.eta = eta
 
         t_last = self.scheduler.timesteps[0] + 1

src/diffusers/pipelines/shap_e/pipeline_shap_e.py

@@ -130,24 +130,6 @@ class ShapEPipeline(DiffusionPipeline):
         latents = latents * scheduler.init_noise_sigma
         return latents
 
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [self.text_encoder, self.prior]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
@@ -176,24 +158,6 @@ class ShapEPipeline(DiffusionPipeline):
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if self.device != torch.device("meta") or not hasattr(self.text_encoder, "_hf_hook"):
-            return self.device
-        for module in self.text_encoder.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     def _encode_prompt(
         self,
         prompt,

src/diffusers/pipelines/shap_e/pipeline_shap_e_img2img.py

@@ -25,7 +25,6 @@ from ...pipelines import DiffusionPipeline
 from ...schedulers import HeunDiscreteScheduler
 from ...utils import (
     BaseOutput,
-    is_accelerate_available,
     logging,
     randn_tensor,
     replace_example_docstring,
@@ -130,42 +129,6 @@ class ShapEImg2ImgPipeline(DiffusionPipeline):
         latents = latents * scheduler.init_noise_sigma
         return latents
 
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
-        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
-        when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        models = [self.image_encoder, self.prior]
-        for cpu_offloaded_model in models:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
-    @property
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if self.device != torch.device("meta") or not hasattr(self.image_encoder, "_hf_hook"):
-            return self.device
-        for module in self.image_encoder.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     def _encode_image(
         self,
         image,

src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py

@@ -231,32 +231,6 @@ class CycleDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lor
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
-        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
-        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
-        Note that offloading happens on a submodule basis. Memory savings are higher than with
-        `enable_model_cpu_offload`, but performance is lower.
-        """
-        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        if self.device.type != "cpu":
-            self.to("cpu", silence_dtype_warnings=True)
-            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
-
-        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
-            cpu_offload(cpu_offloaded_model, device)
-
-        if self.safety_checker is not None:
-            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
@@ -286,25 +260,6 @@ class CycleDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lor
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
     def _encode_prompt(
         self,

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -228,31 +228,6 @@ class StableDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lo
         """
         self.vae.disable_tiling()
 
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
-        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
-        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
-        Note that offloading happens on a submodule basis. Memory savings are higher than with
-        `enable_model_cpu_offload`, but performance is lower.
-        """
-        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        if self.device.type != "cpu":
-            self.to("cpu", silence_dtype_warnings=True)
-            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
-
-        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
-            cpu_offload(cpu_offloaded_model, device)
-
-        if self.safety_checker is not None:
-            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
-
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
@@ -281,24 +256,6 @@ class StableDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lo
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     def _encode_prompt(
         self,
         prompt,

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_attend_and_excite.py

@@ -27,7 +27,7 @@ from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...models.attention_processor import Attention
 from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import is_accelerate_available, is_accelerate_version, logging, randn_tensor, replace_example_docstring
+from ...utils import logging, randn_tensor, replace_example_docstring
 from ..pipeline_utils import DiffusionPipeline
 from . import StableDiffusionPipelineOutput
 from .safety_checker import StableDiffusionSafetyChecker
@@ -251,51 +251,6 @@ class StableDiffusionAttendAndExcitePipeline(DiffusionPipeline, TextualInversion
         """
         self.vae.disable_slicing()
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
-        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
-        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
-        Note that offloading happens on a submodule basis. Memory savings are higher than with
-        `enable_model_cpu_offload`, but performance is lower.
-        """
-        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        if self.device.type != "cpu":
-            self.to("cpu", silence_dtype_warnings=True)
-            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
-
-        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
-            cpu_offload(cpu_offloaded_model, device)
-
-        if self.safety_checker is not None:
-            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
-
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
     def _encode_prompt(
         self,

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py

@@ -28,7 +28,7 @@ from ...image_processor import VaeImageProcessor
 from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import PIL_INTERPOLATION, deprecate, is_accelerate_available, logging, randn_tensor
+from ...utils import PIL_INTERPOLATION, deprecate, logging, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
 
 
@@ -137,42 +137,6 @@ class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoader
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
 
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
-        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
-        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.depth_estimator]:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
     def _encode_prompt(
         self,

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_diffedit.py

@@ -403,32 +403,6 @@ class StableDiffusionDiffEditPipeline(DiffusionPipeline, TextualInversionLoaderM
         """
         self.vae.disable_tiling()
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
-        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
-        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
-        Note that offloading happens on a submodule basis. Memory savings are higher than with
-        `enable_model_cpu_offload`, but performance is lower.
-        """
-        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        if self.device.type != "cpu":
-            self.to("cpu", silence_dtype_warnings=True)
-            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
-
-        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
-            cpu_offload(cpu_offloaded_model, device)
-
-        if self.safety_checker is not None:
-            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
@@ -458,25 +432,6 @@ class StableDiffusionDiffEditPipeline(DiffusionPipeline, TextualInversionLoaderM
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
     def _encode_prompt(
         self,

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py

@@ -25,7 +25,7 @@ from ...configuration_utils import FrozenDict
 from ...image_processor import VaeImageProcessor
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import deprecate, is_accelerate_available, logging, randn_tensor
+from ...utils import deprecate, logging, randn_tensor
 from ..pipeline_utils import DiffusionPipeline
 from . import StableDiffusionPipelineOutput
 from .safety_checker import StableDiffusionSafetyChecker
@@ -123,42 +123,6 @@ class StableDiffusionImageVariationPipeline(DiffusionPipeline):
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
-        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
-        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
-        """
-        if is_accelerate_available():
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        for cpu_offloaded_model in [self.unet, self.image_encoder, self.vae, self.safety_checker]:
-            if cpu_offloaded_model is not None:
-                cpu_offload(cpu_offloaded_model, device)
-
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     def _encode_image(self, image, device, num_images_per_prompt, do_classifier_free_guidance):
         dtype = next(self.image_encoder.parameters()).dtype
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -227,32 +227,6 @@ class StableDiffusionImg2ImgPipeline(
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
-    def enable_sequential_cpu_offload(self, gpu_id=0):
-        r"""
-        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
-        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
-        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
-        Note that offloading happens on a submodule basis. Memory savings are higher than with
-        `enable_model_cpu_offload`, but performance is lower.
-        """
-        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
-            from accelerate import cpu_offload
-        else:
-            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
-
-        device = torch.device(f"cuda:{gpu_id}")
-
-        if self.device.type != "cpu":
-            self.to("cpu", silence_dtype_warnings=True)
-            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
-
-        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
-            cpu_offload(cpu_offloaded_model, device)
-
-        if self.safety_checker is not None:
-            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
@@ -282,25 +256,6 @@ class StableDiffusionImg2ImgPipeline(
         # We'll offload the last model manually.
         self.final_offload_hook = hook
 
-    @property
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
-    def _execution_device(self):
-        r"""
-        Returns the device on which the pipeline's models will be executed. After calling
-        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
-        hooks.
-        """
-        if not hasattr(self.unet, "_hf_hook"):
-            return self.device
-        for module in self.unet.modules():
-            if (
-                hasattr(module, "_hf_hook")
-                and hasattr(module._hf_hook, "execution_device")
-                and module._hf_hook.execution_device is not None
-            ):
-                return torch.device(module._hf_hook.execution_device)
-        return self.device
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
     def _encode_prompt(
         self,