[Refactor] Better align `from_single_file` logic with `from_pretrained` (#7496)

* refactor unet single file loading a bit.

* retrieve the unet from create_diffusers_unet_model_from_ldm

* update

* update

* updae

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* tests

* update

* update

* update

* Update docs/source/en/api/single_file.md
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Update docs/source/en/api/single_file.md
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* Update docs/source/en/api/loaders/single_file.md
Co-authored-by: YiYi Xu <yixu310@gmail.com>

* Update src/diffusers/loaders/single_file.py
Co-authored-by: YiYi Xu <yixu310@gmail.com>

* Update docs/source/en/api/loaders/single_file.md
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Update docs/source/en/api/loaders/single_file.md
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Update docs/source/en/api/loaders/single_file.md
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Update docs/source/en/api/loaders/single_file.md
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

---------
Co-authored-by: sayakpaul <spsayakpaul@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>

tests/single_file/test_stable_diffusion_xl_adapter_single_file.py

+import gc
+import tempfile
+import unittest
+
+import torch
+
+from diffusers import (
+    StableDiffusionXLAdapterPipeline,
+    T2IAdapter,
+)
+from diffusers.utils import load_image
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+)
+
+from .single_file_testing_utils import (
+    SDXLSingleFileTesterMixin,
+    download_diffusers_config,
+    download_original_config,
+    download_single_file_checkpoint,
+)
+
+
+enable_full_determinism()
+
+
+@slow
+@require_torch_gpu
+class StableDiffusionXLAdapterPipelineSingleFileSlowTests(unittest.TestCase, SDXLSingleFileTesterMixin):
+    pipeline_class = StableDiffusionXLAdapterPipeline
+    ckpt_path = "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors"
+    repo_id = "stabilityai/stable-diffusion-xl-base-1.0"
+    original_config = (
+        "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml"
+    )
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_inputs(self):
+        prompt = "toy"
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/t2i_adapter/toy_canny.png"
+        )
+
+        inputs = {
+            "prompt": prompt,
+            "image": image,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 7.5,
+            "output_type": "np",
+        }
+
+        return inputs
+
+    def test_single_file_format_inference_is_same_as_pretrained(self):
+        adapter = T2IAdapter.from_pretrained("TencentARC/t2i-adapter-lineart-sdxl-1.0", torch_dtype=torch.float16)
+        pipe_single_file = StableDiffusionXLAdapterPipeline.from_single_file(
+            self.ckpt_path,
+            adapter=adapter,
+            torch_dtype=torch.float16,
+            safety_checker=None,
+        )
+        pipe_single_file.enable_model_cpu_offload()
+        pipe_single_file.set_progress_bar_config(disable=None)
+
+        inputs = self.get_inputs()
+        images_single_file = pipe_single_file(**inputs).images[0]
+
+        pipe = StableDiffusionXLAdapterPipeline.from_pretrained(
+            self.repo_id,
+            adapter=adapter,
+            torch_dtype=torch.float16,
+            safety_checker=None,
+        )
+        pipe.enable_model_cpu_offload()
+
+        inputs = self.get_inputs()
+        images = pipe(**inputs).images[0]
+
+        assert images_single_file.shape == (768, 512, 3)
+        assert images.shape == (768, 512, 3)
+
+        max_diff = numpy_cosine_similarity_distance(images.flatten(), images_single_file.flatten())
+        assert max_diff < 5e-3
+
+    def test_single_file_components(self):
+        adapter = T2IAdapter.from_pretrained("TencentARC/t2i-adapter-lineart-sdxl-1.0", torch_dtype=torch.float16)
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            adapter=adapter,
+            torch_dtype=torch.float16,
+        )
+
+        pipe_single_file = self.pipeline_class.from_single_file(self.ckpt_path, safety_checker=None, adapter=adapter)
+        super().test_single_file_components(pipe, pipe_single_file)
+
+    def test_single_file_components_local_files_only(self):
+        adapter = T2IAdapter.from_pretrained("TencentARC/t2i-adapter-lineart-sdxl-1.0", torch_dtype=torch.float16)
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            adapter=adapter,
+            torch_dtype=torch.float16,
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            ckpt_filename = self.ckpt_path.split("/")[-1]
+            local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
+
+            single_file_pipe = self.pipeline_class.from_single_file(
+                local_ckpt_path, adapter=adapter, safety_checker=None, local_files_only=True
+            )
+
+        self._compare_component_configs(pipe, single_file_pipe)
+
+    def test_single_file_components_with_diffusers_config(self):
+        adapter = T2IAdapter.from_pretrained("TencentARC/t2i-adapter-lineart-sdxl-1.0", torch_dtype=torch.float16)
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            adapter=adapter,
+            torch_dtype=torch.float16,
+            safety_checker=None,
+        )
+
+        pipe_single_file = self.pipeline_class.from_single_file(self.ckpt_path, config=self.repo_id, adapter=adapter)
+        self._compare_component_configs(pipe, pipe_single_file)
+
+    def test_single_file_components_with_diffusers_config_local_files_only(self):
+        adapter = T2IAdapter.from_pretrained("TencentARC/t2i-adapter-lineart-sdxl-1.0", torch_dtype=torch.float16)
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            adapter=adapter,
+            torch_dtype=torch.float16,
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            ckpt_filename = self.ckpt_path.split("/")[-1]
+            local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
+            local_diffusers_config = download_diffusers_config(self.repo_id, tmpdir)
+
+            pipe_single_file = self.pipeline_class.from_single_file(
+                local_ckpt_path,
+                config=local_diffusers_config,
+                adapter=adapter,
+                safety_checker=None,
+                local_files_only=True,
+            )
+        self._compare_component_configs(pipe, pipe_single_file)
+
+    def test_single_file_components_with_original_config(self):
+        adapter = T2IAdapter.from_pretrained("TencentARC/t2i-adapter-lineart-sdxl-1.0", torch_dtype=torch.float16)
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            adapter=adapter,
+            torch_dtype=torch.float16,
+            safety_checker=None,
+        )
+
+        pipe_single_file = self.pipeline_class.from_single_file(
+            self.ckpt_path, original_config=self.original_config, adapter=adapter
+        )
+        self._compare_component_configs(pipe, pipe_single_file)
+
+    def test_single_file_components_with_original_config_local_files_only(self):
+        adapter = T2IAdapter.from_pretrained("TencentARC/t2i-adapter-lineart-sdxl-1.0", torch_dtype=torch.float16)
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            adapter=adapter,
+            torch_dtype=torch.float16,
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            ckpt_filename = self.ckpt_path.split("/")[-1]
+            local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
+            local_original_config = download_original_config(self.original_config, tmpdir)
+
+            pipe_single_file = self.pipeline_class.from_single_file(
+                local_ckpt_path,
+                original_config=local_original_config,
+                adapter=adapter,
+                safety_checker=None,
+                local_files_only=True,
+            )
+        self._compare_component_configs(pipe, pipe_single_file)

tests/single_file/test_stable_diffusion_xl_controlnet_single_file.py

+import gc
+import tempfile
+import unittest
+
+import torch
+
+from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
+from diffusers.utils import load_image
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from .single_file_testing_utils import (
+    SDXLSingleFileTesterMixin,
+    download_diffusers_config,
+    download_single_file_checkpoint,
+)
+
+
+enable_full_determinism()
+
+
+@slow
+@require_torch_gpu
+class StableDiffusionXLControlNetPipelineSingleFileSlowTests(unittest.TestCase, SDXLSingleFileTesterMixin):
+    pipeline_class = StableDiffusionXLControlNetPipeline
+    ckpt_path = "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors"
+    repo_id = "stabilityai/stable-diffusion-xl-base-1.0"
+    original_config = (
+        "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml"
+    )
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
+        generator = torch.Generator(device=generator_device).manual_seed(seed)
+        image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/stormtrooper_depth.png"
+        )
+        inputs = {
+            "prompt": "Stormtrooper's lecture",
+            "image": image,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "strength": 0.75,
+            "guidance_scale": 7.5,
+            "output_type": "np",
+        }
+
+        return inputs
+
+    def test_single_file_format_inference_is_same_as_pretrained(self):
+        controlnet = ControlNetModel.from_pretrained("diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16)
+        pipe_single_file = self.pipeline_class.from_single_file(
+            self.ckpt_path, controlnet=controlnet, torch_dtype=torch.float16
+        )
+        pipe_single_file.unet.set_default_attn_processor()
+        pipe_single_file.enable_model_cpu_offload()
+        pipe_single_file.set_progress_bar_config(disable=None)
+
+        inputs = self.get_inputs(torch_device)
+        single_file_images = pipe_single_file(**inputs).images[0]
+
+        pipe = self.pipeline_class.from_pretrained(self.repo_id, controlnet=controlnet, torch_dtype=torch.float16)
+        pipe.unet.set_default_attn_processor()
+        pipe.enable_model_cpu_offload()
+
+        inputs = self.get_inputs(torch_device)
+        images = pipe(**inputs).images[0]
+
+        assert images.shape == (512, 512, 3)
+        assert single_file_images.shape == (512, 512, 3)
+
+        max_diff = numpy_cosine_similarity_distance(images[0].flatten(), single_file_images[0].flatten())
+        assert max_diff < 5e-2
+
+    def test_single_file_components(self):
+        controlnet = ControlNetModel.from_pretrained(
+            "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16"
+        )
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            controlnet=controlnet,
+            torch_dtype=torch.float16,
+        )
+
+        pipe_single_file = self.pipeline_class.from_single_file(self.ckpt_path, controlnet=controlnet)
+        super().test_single_file_components(pipe, pipe_single_file)
+
+    def test_single_file_components_local_files_only(self):
+        controlnet = ControlNetModel.from_pretrained(
+            "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16"
+        )
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            controlnet=controlnet,
+            torch_dtype=torch.float16,
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            ckpt_filename = self.ckpt_path.split("/")[-1]
+            local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
+
+            single_file_pipe = self.pipeline_class.from_single_file(
+                local_ckpt_path, controlnet=controlnet, safety_checker=None, local_files_only=True
+            )
+
+        self._compare_component_configs(pipe, single_file_pipe)
+
+    def test_single_file_components_with_original_config(self):
+        controlnet = ControlNetModel.from_pretrained(
+            "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16"
+        )
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            controlnet=controlnet,
+            torch_dtype=torch.float16,
+        )
+
+        pipe_single_file = self.pipeline_class.from_single_file(
+            self.ckpt_path,
+            original_config=self.original_config,
+            controlnet=controlnet,
+        )
+        self._compare_component_configs(pipe, pipe_single_file)
+
+    def test_single_file_components_with_original_config_local_files_only(self):
+        controlnet = ControlNetModel.from_pretrained(
+            "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16"
+        )
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            variant="fp16",
+            controlnet=controlnet,
+            torch_dtype=torch.float16,
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            ckpt_filename = self.ckpt_path.split("/")[-1]
+            local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
+
+            pipe_single_file = self.pipeline_class.from_single_file(
+                local_ckpt_path,
+                safety_checker=None,
+                controlnet=controlnet,
+                local_files_only=True,
+            )
+        self._compare_component_configs(pipe, pipe_single_file)
+
+    def test_single_file_components_with_diffusers_config(self):
+        controlnet = ControlNetModel.from_pretrained(
+            "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16"
+        )
+        pipe = self.pipeline_class.from_pretrained(self.repo_id, controlnet=controlnet)
+        pipe_single_file = self.pipeline_class.from_single_file(
+            self.ckpt_path, controlnet=controlnet, config=self.repo_id
+        )
+
+        super()._compare_component_configs(pipe, pipe_single_file)
+
+    def test_single_file_components_with_diffusers_config_local_files_only(self):
+        controlnet = ControlNetModel.from_pretrained(
+            "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16"
+        )
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id,
+            controlnet=controlnet,
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            ckpt_filename = self.ckpt_path.split("/")[-1]
+            local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
+            local_diffusers_config = download_diffusers_config(self.repo_id, tmpdir)
+
+            pipe_single_file = self.pipeline_class.from_single_file(
+                local_ckpt_path,
+                config=local_diffusers_config,
+                safety_checker=None,
+                controlnet=controlnet,
+                local_files_only=True,
+            )
+        super()._compare_component_configs(pipe, pipe_single_file)

tests/single_file/test_stable_diffusion_xl_img2img_single_file.py

+import gc
+import unittest
+
+import torch
+
+from diffusers import (
+    DDIMScheduler,
+    StableDiffusionXLImg2ImgPipeline,
+)
+from diffusers.utils import load_image
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+)
+
+from .single_file_testing_utils import SDXLSingleFileTesterMixin
+
+
+enable_full_determinism()
+
+
+@slow
+@require_torch_gpu
+class StableDiffusionXLImg2ImgPipelineSingleFileSlowTests(unittest.TestCase, SDXLSingleFileTesterMixin):
+    pipeline_class = StableDiffusionXLImg2ImgPipeline
+    ckpt_path = "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors"
+    repo_id = "stabilityai/stable-diffusion-xl-base-1.0"
+    original_config = (
+        "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml"
+    )
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
+        generator = torch.Generator(device=generator_device).manual_seed(seed)
+        init_image = load_image(
+            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
+            "/stable_diffusion_img2img/sketch-mountains-input.png"
+        )
+        inputs = {
+            "prompt": "a fantasy landscape, concept art, high resolution",
+            "image": init_image,
+            "generator": generator,
+            "num_inference_steps": 3,
+            "strength": 0.75,
+            "guidance_scale": 7.5,
+            "output_type": "np",
+        }
+        return inputs
+
+    def test_single_file_format_inference_is_same_as_pretrained(self):
+        super().test_single_file_format_inference_is_same_as_pretrained(expected_max_diff=1e-3)
+
+
+@slow
+@require_torch_gpu
+class StableDiffusionXLImg2ImgRefinerPipelineSingleFileSlowTests(unittest.TestCase):
+    pipeline_class = StableDiffusionXLImg2ImgPipeline
+    ckpt_path = (
+        "https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/blob/main/sd_xl_refiner_1.0.safetensors"
+    )
+    repo_id = "stabilityai/stable-diffusion-xl-refiner-1.0"
+    original_config = (
+        "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_refiner.yaml"
+    )
+
+    def test_single_file_format_inference_is_same_as_pretrained(self):
+        init_image = load_image(
+            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
+            "/stable_diffusion_img2img/sketch-mountains-input.png"
+        )
+
+        pipe = self.pipeline_class.from_pretrained(self.repo_id, torch_dtype=torch.float16)
+        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+        pipe.unet.set_default_attn_processor()
+        pipe.enable_model_cpu_offload()
+
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        image = pipe(
+            prompt="mountains", image=init_image, num_inference_steps=5, generator=generator, output_type="np"
+        ).images[0]
+
+        pipe_single_file = self.pipeline_class.from_single_file(self.ckpt_path, torch_dtype=torch.float16)
+        pipe_single_file.scheduler = DDIMScheduler.from_config(pipe_single_file.scheduler.config)
+        pipe_single_file.unet.set_default_attn_processor()
+        pipe_single_file.enable_model_cpu_offload()
+
+        generator = torch.Generator(device="cpu").manual_seed(0)
+        image_single_file = pipe_single_file(
+            prompt="mountains", image=init_image, num_inference_steps=5, generator=generator, output_type="np"
+        ).images[0]
+
+        max_diff = numpy_cosine_similarity_distance(image.flatten(), image_single_file.flatten())
+
+        assert max_diff < 5e-4

tests/single_file/test_stable_diffusion_xl_instruct_pix2pix.py

+import gc
+import unittest
+
+import torch
+
+from diffusers import StableDiffusionXLInstructPix2PixPipeline
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    require_torch_gpu,
+    slow,
+)
+
+
+enable_full_determinism()
+
+
+@slow
+@require_torch_gpu
+class StableDiffusionXLInstructPix2PixPipeline(unittest.TestCase):
+    pipeline_class = StableDiffusionXLInstructPix2PixPipeline
+    ckpt_path = "https://huggingface.co/stabilityai/cosxl/blob/main/cosxl_edit.safetensors"
+    original_config = None
+    repo_id = "diffusers/sdxl-instructpix2pix-768"
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
+        generator = torch.Generator(device=generator_device).manual_seed(seed)
+        inputs = {
+            "prompt": "a fantasy landscape, concept art, high resolution",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "strength": 0.75,
+            "guidance_scale": 7.5,
+            "output_type": "np",
+        }
+        return inputs
+
+    def test_single_file_setting_cosxl_edit(self):
+        # Default is PNDM for this checkpoint
+        pipe = self.pipeline_class.from_single_file(self.ckpt_path, config=self.repo_id, is_cosxl_edit=True)
+        assert pipe.is_cosxl_edit is True

tests/single_file/test_stable_diffusion_xl_single_file.py

+import gc
+import unittest
+
+import torch
+
+from diffusers import (
+    StableDiffusionXLPipeline,
+)
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    require_torch_gpu,
+    slow,
+)
+
+from .single_file_testing_utils import SDXLSingleFileTesterMixin
+
+
+enable_full_determinism()
+
+
+@slow
+@require_torch_gpu
+class StableDiffusionXLPipelineSingleFileSlowTests(unittest.TestCase, SDXLSingleFileTesterMixin):
+    pipeline_class = StableDiffusionXLPipeline
+    ckpt_path = "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors"
+    repo_id = "stabilityai/stable-diffusion-xl-base-1.0"
+    original_config = (
+        "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml"
+    )
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
+        generator = torch.Generator(device=generator_device).manual_seed(seed)
+        inputs = {
+            "prompt": "a fantasy landscape, concept art, high resolution",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "strength": 0.75,
+            "guidance_scale": 7.5,
+            "output_type": "np",
+        }
+        return inputs
+
+    def test_single_file_format_inference_is_same_as_pretrained(self):
+        super().test_single_file_format_inference_is_same_as_pretrained(expected_max_diff=1e-3)