[Core] better support offloading when side loading is enabled. (#4855)

* better support offloading when side loading is enabled.

* load_textual_inversion

* better messaging for textual inversion.

* fixes

* address PR feedback.

* sdxl support.

* improve messaging

* recursive removal when cpu sequential offloading is enabled.

* add: lora tests

* recruse.

* add: offload tests for textual inversion.

src/diffusers/loaders.py

@@ -45,6 +45,7 @@ if is_transformers_available():
 
 if is_accelerate_available():
     from accelerate import init_empty_weights
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
     from accelerate.utils import set_module_tensor_to_device
 
 logger = logging.get_logger(__name__)
@@ -768,6 +769,21 @@ class TextualInversionLoaderMixin:
                 f" `{self.load_textual_inversion.__name__}`"
             )
 
+        # Remove any existing hooks.
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recursive = False
+        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)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    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."
+                    )
+                    recursive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recursive)
+
         cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
@@ -921,6 +937,12 @@ class TextualInversionLoaderMixin:
         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()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
 
 class LoraLoaderMixin:
     r"""
@@ -952,6 +974,21 @@ class LoraLoaderMixin:
             kwargs (`dict`, *optional*):
                 See [`~loaders.LoraLoaderMixin.lora_state_dict`].
         """
+        # Remove any existing hooks.
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recurive = False
+        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)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    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."
+                    )
+                    recurive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recurive)
+
         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(
@@ -961,6 +998,12 @@ class LoraLoaderMixin:
             lora_scale=self.lora_scale,
         )
 
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
     @classmethod
     def lora_state_dict(
         cls,

src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py

@@ -1549,6 +1549,26 @@ class StableDiffusionXLControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMi
         # We could have accessed the unet config from `lora_state_dict()` too. We pass
         # it here explicitly to be able to tell that it's coming from an SDXL
         # pipeline.
+
+        # Remove any existing hooks.
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+        else:
+            raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
+
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recursive = False
+        for _, component in self.components.items():
+            if isinstance(component, torch.nn.Module):
+                if hasattr(component, "_hf_hook"):
+                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    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."
+                    )
+                    recursive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recursive)
         state_dict, network_alphas = self.lora_state_dict(
             pretrained_model_name_or_path_or_dict,
             unet_config=self.unet.config,
@@ -1576,6 +1596,12 @@ class StableDiffusionXLControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMi
                 lora_scale=self.lora_scale,
             )
 
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
     @classmethod
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.save_lora_weights
     def save_lora_weights(

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py

@@ -1212,6 +1212,26 @@ class StableDiffusionXLControlNetPipeline(
         # We could have accessed the unet config from `lora_state_dict()` too. We pass
         # it here explicitly to be able to tell that it's coming from an SDXL
         # pipeline.
+
+        # Remove any existing hooks.
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+        else:
+            raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
+
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recursive = False
+        for _, component in self.components.items():
+            if isinstance(component, torch.nn.Module):
+                if hasattr(component, "_hf_hook"):
+                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    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."
+                    )
+                    recursive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recursive)
         state_dict, network_alphas = self.lora_state_dict(
             pretrained_model_name_or_path_or_dict,
             unet_config=self.unet.config,
@@ -1239,6 +1259,12 @@ class StableDiffusionXLControlNetPipeline(
                 lora_scale=self.lora_scale,
             )
 
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
     @classmethod
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.save_lora_weights
     def save_lora_weights(

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

@@ -916,6 +916,26 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoad
         # We could have accessed the unet config from `lora_state_dict()` too. We pass
         # it here explicitly to be able to tell that it's coming from an SDXL
         # pipeline.
+
+        # Remove any existing hooks.
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+        else:
+            raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
+
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recursive = False
+        for _, component in self.components.items():
+            if isinstance(component, torch.nn.Module):
+                if hasattr(component, "_hf_hook"):
+                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    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."
+                    )
+                    recursive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recursive)
         state_dict, network_alphas = self.lora_state_dict(
             pretrained_model_name_or_path_or_dict,
             unet_config=self.unet.config,
@@ -943,6 +963,12 @@ class StableDiffusionXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoad
                 lora_scale=self.lora_scale,
             )
 
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
     @classmethod
     def save_lora_weights(
         self,

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -1070,6 +1070,26 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin, L
         # We could have accessed the unet config from `lora_state_dict()` too. We pass
         # it here explicitly to be able to tell that it's coming from an SDXL
         # pipeline.
+
+        # Remove any existing hooks.
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+        else:
+            raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
+
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recursive = False
+        for _, component in self.components.items():
+            if isinstance(component, torch.nn.Module):
+                if hasattr(component, "_hf_hook"):
+                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    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."
+                    )
+                    recursive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recursive)
         state_dict, network_alphas = self.lora_state_dict(
             pretrained_model_name_or_path_or_dict,
             unet_config=self.unet.config,
@@ -1097,6 +1117,12 @@ class StableDiffusionXLImg2ImgPipeline(DiffusionPipeline, FromSingleFileMixin, L
                 lora_scale=self.lora_scale,
             )
 
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
     @classmethod
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.save_lora_weights
     def save_lora_weights(

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py

@@ -1384,6 +1384,26 @@ class StableDiffusionXLInpaintPipeline(DiffusionPipeline, LoraLoaderMixin, FromS
         # We could have accessed the unet config from `lora_state_dict()` too. We pass
         # it here explicitly to be able to tell that it's coming from an SDXL
         # pipeline.
+
+        # Remove any existing hooks.
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+        else:
+            raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
+
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recursive = False
+        for _, component in self.components.items():
+            if isinstance(component, torch.nn.Module):
+                if hasattr(component, "_hf_hook"):
+                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    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."
+                    )
+                    recursive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recursive)
         state_dict, network_alphas = self.lora_state_dict(
             pretrained_model_name_or_path_or_dict,
             unet_config=self.unet.config,
@@ -1411,6 +1431,12 @@ class StableDiffusionXLInpaintPipeline(DiffusionPipeline, LoraLoaderMixin, FromS
                 lora_scale=self.lora_scale,
             )
 
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
     @classmethod
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.save_lora_weights
     def save_lora_weights(

tests/models/test_lora_layers.py

@@ -1081,6 +1081,42 @@ class LoraIntegrationTests(unittest.TestCase):
 
         self.assertTrue(np.allclose(images, expected, atol=1e-3))
 
+    def test_a1111_with_model_cpu_offload(self):
+        generator = torch.Generator().manual_seed(0)
+
+        pipe = StableDiffusionPipeline.from_pretrained("hf-internal-testing/Counterfeit-V2.5", safety_checker=None)
+        pipe.enable_model_cpu_offload()
+        lora_model_id = "hf-internal-testing/civitai-light-shadow-lora"
+        lora_filename = "light_and_shadow.safetensors"
+        pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
+
+        images = pipe(
+            "masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
+        ).images
+
+        images = images[0, -3:, -3:, -1].flatten()
+        expected = np.array([0.3636, 0.3708, 0.3694, 0.3679, 0.3829, 0.3677, 0.3692, 0.3688, 0.3292])
+
+        self.assertTrue(np.allclose(images, expected, atol=1e-3))
+
+    def test_a1111_with_sequential_cpu_offload(self):
+        generator = torch.Generator().manual_seed(0)
+
+        pipe = StableDiffusionPipeline.from_pretrained("hf-internal-testing/Counterfeit-V2.5", safety_checker=None)
+        pipe.enable_sequential_cpu_offload()
+        lora_model_id = "hf-internal-testing/civitai-light-shadow-lora"
+        lora_filename = "light_and_shadow.safetensors"
+        pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
+
+        images = pipe(
+            "masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
+        ).images
+
+        images = images[0, -3:, -3:, -1].flatten()
+        expected = np.array([0.3636, 0.3708, 0.3694, 0.3679, 0.3829, 0.3677, 0.3692, 0.3688, 0.3292])
+
+        self.assertTrue(np.allclose(images, expected, atol=1e-3))
+
     def test_kohya_sd_v15_with_higher_dimensions(self):
         generator = torch.Generator().manual_seed(0)
 
@@ -1257,10 +1293,10 @@ class LoraIntegrationTests(unittest.TestCase):
         generator = torch.Generator().manual_seed(0)
 
         pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")
+        pipe.enable_model_cpu_offload()
         lora_model_id = "hf-internal-testing/sdxl-1.0-lora"
         lora_filename = "sd_xl_offset_example-lora_1.0.safetensors"
         pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
-        pipe.enable_model_cpu_offload()
 
         images = pipe(
             "masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
@@ -1413,3 +1449,21 @@ class LoraIntegrationTests(unittest.TestCase):
         assert state_dicts_almost_equal(text_encoder_1_sd, new_text_encoder_1_sd)
         assert state_dicts_almost_equal(text_encoder_2_sd, new_text_encoder_2_sd)
         assert state_dicts_almost_equal(unet_sd, new_unet_sd)
+
+    def test_sdxl_1_0_lora_with_sequential_cpu_offloading(self):
+        generator = torch.Generator().manual_seed(0)
+
+        pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")
+        pipe.enable_sequential_cpu_offload()
+        lora_model_id = "hf-internal-testing/sdxl-1.0-lora"
+        lora_filename = "sd_xl_offset_example-lora_1.0.safetensors"
+        pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
+
+        images = pipe(
+            "masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
+        ).images
+
+        images = images[0, -3:, -3:, -1].flatten()
+        expected = np.array([0.4468, 0.4087, 0.4134, 0.366, 0.3202, 0.3505, 0.3786, 0.387, 0.3535])
+
+        self.assertTrue(np.allclose(images, expected, atol=1e-3))

tests/pipelines/stable_diffusion/test_stable_diffusion.py

@@ -1019,6 +1019,56 @@ class StableDiffusionPipelineSlowTests(unittest.TestCase):
         max_diff = np.abs(expected_image - image).max()
         assert max_diff < 8e-1
 
+    def test_stable_diffusion_textual_inversion_with_model_cpu_offload(self):
+        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+        pipe.enable_model_cpu_offload()
+        pipe.load_textual_inversion("sd-concepts-library/low-poly-hd-logos-icons")
+
+        a111_file = hf_hub_download("hf-internal-testing/text_inv_embedding_a1111_format", "winter_style.pt")
+        a111_file_neg = hf_hub_download(
+            "hf-internal-testing/text_inv_embedding_a1111_format", "winter_style_negative.pt"
+        )
+        pipe.load_textual_inversion(a111_file)
+        pipe.load_textual_inversion(a111_file_neg)
+
+        generator = torch.Generator(device="cpu").manual_seed(1)
+
+        prompt = "An logo of a turtle in strong Style-Winter with <low-poly-hd-logos-icons>"
+        neg_prompt = "Style-Winter-neg"
+
+        image = pipe(prompt=prompt, negative_prompt=neg_prompt, generator=generator, output_type="np").images[0]
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/text_inv/winter_logo_style.npy"
+        )
+
+        max_diff = np.abs(expected_image - image).max()
+        assert max_diff < 8e-1
+
+    def test_stable_diffusion_textual_inversion_with_sequential_cpu_offload(self):
+        pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+        pipe.enable_sequential_cpu_offload()
+        pipe.load_textual_inversion("sd-concepts-library/low-poly-hd-logos-icons")
+
+        a111_file = hf_hub_download("hf-internal-testing/text_inv_embedding_a1111_format", "winter_style.pt")
+        a111_file_neg = hf_hub_download(
+            "hf-internal-testing/text_inv_embedding_a1111_format", "winter_style_negative.pt"
+        )
+        pipe.load_textual_inversion(a111_file)
+        pipe.load_textual_inversion(a111_file_neg)
+
+        generator = torch.Generator(device="cpu").manual_seed(1)
+
+        prompt = "An logo of a turtle in strong Style-Winter with <low-poly-hd-logos-icons>"
+        neg_prompt = "Style-Winter-neg"
+
+        image = pipe(prompt=prompt, negative_prompt=neg_prompt, generator=generator, output_type="np").images[0]
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/text_inv/winter_logo_style.npy"
+        )
+
+        max_diff = np.abs(expected_image - image).max()
+        assert max_diff < 8e-1
+
     @require_torch_2
     def test_stable_diffusion_compile(self):
         seed = 0