pinn ruff

.github/workflows/pr_test_fetcher.yml

@@ -1,6 +1,12 @@
 name: Fast tests for PRs - Test Fetcher
 
-on: workflow_dispatch
+on:
+  pull_request:
+    branches:
+      - main
+  push:
+    branches:
+      - ci-*
 
 env:
   DIFFUSERS_IS_CI: yes

.github/workflows/pr_tests.yml

@@ -113,7 +113,6 @@ jobs:
     - name: Run example PyTorch CPU tests
       if: ${{ matrix.config.framework == 'pytorch_examples' }}
       run: |
-        python -m pip install peft
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           --make-reports=tests_${{ matrix.config.report }} \
           examples

docs/source/en/_toctree.yml

@@ -264,10 +264,6 @@
       title: ControlNet
     - local: api/pipelines/controlnet_sdxl
       title: ControlNet with Stable Diffusion XL
-    - local: api/pipelines/controlnetxs
-      title: ControlNet-XS
-    - local: api/pipelines/controlnetxs_sdxl
-      title: ControlNet-XS with Stable Diffusion XL
     - local: api/pipelines/cycle_diffusion
       title: Cycle Diffusion
     - local: api/pipelines/dance_diffusion

docs/source/en/api/attnprocessor.md

@@ -20,9 +20,6 @@ An attention processor is a class for applying different types of attention mech
 ## AttnProcessor2_0
 [[autodoc]] models.attention_processor.AttnProcessor2_0
 
-## FusedAttnProcessor2_0
-[[autodoc]] models.attention_processor.FusedAttnProcessor2_0
-
 ## LoRAAttnProcessor
 [[autodoc]] models.attention_processor.LoRAAttnProcessor
 

docs/source/en/api/pipelines/controlnetxs.md

@@ -1,39 +0,0 @@
-<!--Copyright 2023 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
--->
-
-# ControlNet-XS
-
-ControlNet-XS was introduced in [ControlNet-XS](https://vislearn.github.io/ControlNet-XS/) by Denis Zavadski and Carsten Rother. It is based on the observation that the control model in the [original ControlNet](https://huggingface.co/papers/2302.05543) can be made much smaller and still produce good results.
-
-Like the original ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
-
-ControlNet-XS generates images with comparable quality to a regular ControlNet, but it is 20-25% faster ([see benchmark](https://github.com/UmerHA/controlnet-xs-benchmark/blob/main/Speed%20Benchmark.ipynb) with StableDiffusion-XL) and uses ~45% less memory.
-
-Here's the overview from the [project page](https://vislearn.github.io/ControlNet-XS/):
-
-*With increasing computing capabilities, current model architectures appear to follow the trend of simply upscaling all components without validating the necessity for doing so. In this project we investigate the size and architectural design of ControlNet [Zhang et al., 2023] for controlling the image generation process with stable diffusion-based models. We show that a new architecture with as little as 1% of the parameters of the base model achieves state-of-the art results, considerably better than ControlNet in terms of FID score. Hence we call it ControlNet-XS. We provide the code for controlling StableDiffusion-XL [Podell et al., 2023] (Model B, 48M Parameters) and StableDiffusion 2.1 [Rombach et al. 2022] (Model B, 14M Parameters), all under openrail license.*
-
-This model was contributed by [UmerHA](https://twitter.com/UmerHAdil). ❤️
-
-<Tip>
-
-Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-components-across-pipelines) section to learn how to efficiently load the same components into multiple pipelines.
-
-</Tip>
-
-## StableDiffusionControlNetXSPipeline
-[[autodoc]] StableDiffusionControlNetXSPipeline
-	- all
-	- __call__
-
-## StableDiffusionPipelineOutput
-[[autodoc]] pipelines.stable_diffusion.StableDiffusionPipelineOutput

docs/source/en/api/pipelines/controlnetxs_sdxl.md

@@ -1,45 +0,0 @@
-<!--Copyright 2023 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
--->
-
-# ControlNet-XS with Stable Diffusion XL
-
-ControlNet-XS was introduced in [ControlNet-XS](https://vislearn.github.io/ControlNet-XS/) by Denis Zavadski and Carsten Rother. It is based on the observation that the control model in the [original ControlNet](https://huggingface.co/papers/2302.05543) can be made much smaller and still produce good results.
-
-Like the original ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
-
-ControlNet-XS generates images with comparable quality to a regular ControlNet, but it is 20-25% faster ([see benchmark](https://github.com/UmerHA/controlnet-xs-benchmark/blob/main/Speed%20Benchmark.ipynb)) and uses ~45% less memory.
-
-Here's the overview from the [project page](https://vislearn.github.io/ControlNet-XS/):
-
-*With increasing computing capabilities, current model architectures appear to follow the trend of simply upscaling all components without validating the necessity for doing so. In this project we investigate the size and architectural design of ControlNet [Zhang et al., 2023] for controlling the image generation process with stable diffusion-based models. We show that a new architecture with as little as 1% of the parameters of the base model achieves state-of-the art results, considerably better than ControlNet in terms of FID score. Hence we call it ControlNet-XS. We provide the code for controlling StableDiffusion-XL [Podell et al., 2023] (Model B, 48M Parameters) and StableDiffusion 2.1 [Rombach et al. 2022] (Model B, 14M Parameters), all under openrail license.*
-
-This model was contributed by [UmerHA](https://twitter.com/UmerHAdil). ❤️
-
-<Tip warning={true}>
-
-🧪 Many of the SDXL ControlNet checkpoints are experimental, and there is a lot of room for improvement. Feel free to open an [Issue](https://github.com/huggingface/diffusers/issues/new/choose) and leave us feedback on how we can improve!
-
-</Tip>
-
-<Tip>
-
-Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-components-across-pipelines) section to learn how to efficiently load the same components into multiple pipelines.
-
-</Tip>
-
-## StableDiffusionXLControlNetXSPipeline
-[[autodoc]] StableDiffusionXLControlNetXSPipeline
-	- all
-	- __call__
-
-## StableDiffusionPipelineOutput
-[[autodoc]] pipelines.stable_diffusion.StableDiffusionPipelineOutput

docs/source/en/api/pipelines/overview.md

@@ -40,8 +40,6 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [Consistency Models](consistency_models) | unconditional image generation |
 | [ControlNet](controlnet) | text2image, image2image, inpainting |
 | [ControlNet with Stable Diffusion XL](controlnet_sdxl) | text2image |
-| [ControlNet-XS](controlnetxs) | text2image |
-| [ControlNet-XS with Stable Diffusion XL](controlnetxs_sdxl) | text2image |
 | [Cycle Diffusion](cycle_diffusion) | image2image |
 | [Dance Diffusion](dance_diffusion) | unconditional audio generation |
 | [DDIM](ddim) | unconditional image generation |
@@ -73,7 +71,6 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [Stable Diffusion](stable_diffusion/overview) | text2image, image2image, depth2image, inpainting, image variation, latent upscaler, super-resolution |
 | [Stable Diffusion Model Editing](model_editing) | model editing |
 | [Stable Diffusion XL](stable_diffusion/stable_diffusion_xl) | text2image, image2image, inpainting |
-| [Stable Diffusion XL Turbo](stable_diffusion/sdxl_turbo) | text2image, image2image, inpainting |
 | [Stable unCLIP](stable_unclip) | text2image, image variation |
 | [Stochastic Karras VE](stochastic_karras_ve) | unconditional image generation |
 | [T2I-Adapter](stable_diffusion/adapter) | text2image |

docs/source/en/api/pipelines/stable_diffusion/sdxl_turbo.md

@@ -20,7 +20,7 @@ The abstract from the paper is:
 
 ## Tips
 
-- SDXL Turbo uses the exact same architecture as [SDXL](./stable_diffusion_xl), which means it also has the same API. Please refer to the [SDXL](./stable_diffusion_xl) API reference for more details.
+- SDXL Turbo uses the exact same architecture as [SDXL](./stable_diffusion_xl).
 - SDXL Turbo should disable guidance scale by setting `guidance_scale=0.0`
 - SDXL Turbo should use `timestep_spacing='trailing'` for the scheduler and use between 1 and 4 steps.
 - SDXL Turbo has been trained to generate images of size 512x512.
@@ -28,8 +28,26 @@ The abstract from the paper is:
 
 <Tip>
 
-To learn how to use SDXL Turbo for various tasks, how to optimize performance, and other usage examples, take a look at the [SDXL Turbo](../../../using-diffusers/sdxl_turbo) guide.
+To learn how to use SDXL Turbo for various tasks, how to optimize performance, and other usage examples, take a look at the [Stable Diffusion XL](../../../using-diffusers/sdxl_turbo) guide.
 
 Check out the [Stability AI](https://huggingface.co/stabilityai) Hub organization for the official base and refiner model checkpoints!
 
 </Tip>
+
+## StableDiffusionXLPipeline
+
+[[autodoc]] StableDiffusionXLPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLImg2ImgPipeline
+
+[[autodoc]] StableDiffusionXLImg2ImgPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLInpaintPipeline
+
+[[autodoc]] StableDiffusionXLInpaintPipeline
+	- all
+	- __call__

docs/source/en/using-diffusers/loading_adapters.md

@@ -485,69 +485,6 @@ image.save("sdxl_t2i.png")
   </div>
 </div>
 
-You can use the IP-Adapter face model to apply specific faces to your images.  It is an effective way to maintain consistent characters in your image generations.
-Weights are loaded with the same method used for the other IP-Adapters.  
-
-```python
-# Load ip-adapter-full-face_sd15.bin
-pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter-full-face_sd15.bin")
-```
-
-<Tip>
-
-It is recommended to use `DDIMScheduler` and `EulerDiscreteScheduler` for face model. 
-
-
-</Tip>
-
-```python
-import torch
-from diffusers import StableDiffusionPipeline, DDIMScheduler
-from diffusers.utils import load_image
-
-noise_scheduler = DDIMScheduler(
-    num_train_timesteps=1000,
-    beta_start=0.00085,
-    beta_end=0.012,
-    beta_schedule="scaled_linear",
-    clip_sample=False,
-    set_alpha_to_one=False,
-    steps_offset=1
-)
-
-pipeline = StableDiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-v1-5",
-    torch_dtype=torch.float16,
-    scheduler=noise_scheduler,
-).to("cuda")
-
-pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter-full-face_sd15.bin")
-
-pipeline.set_ip_adapter_scale(0.7)
-
-image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ai_face2.png")
-
-generator = torch.Generator(device="cpu").manual_seed(33)
-
-image = pipeline(
-    prompt="A photo of a girl wearing a black dress, holding red roses in hand, upper body, behind is the Eiffel Tower",
-    ip_adapter_image=image,
-    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
-    num_inference_steps=50, num_images_per_prompt=1, width=512, height=704,
-    generator=generator,
-).images[0]
-```
-
-<div class="flex flex-row gap-4">
-  <div class="flex-1">
-    <img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ai_face2.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">input image</figcaption>
-  </div>
-  <div class="flex-1">
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ipadapter_full_face_output.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">output image</figcaption>
-  </div>
-</div>
 
 ### LCM-Lora
 

docs/source/en/using-diffusers/push_to_hub.md

@@ -174,4 +174,10 @@ Set `private=True` in the [`~diffusers.utils.PushToHubMixin.push_to_hub`] functi
 controlnet.push_to_hub("my-controlnet-model-private", private=True)
 ```
 
-Private repositories are only visible to you, and other users won't be able to clone the repository and your repository won't appear in search results. Even if a user has the URL to your private repository, they'll receive a `404 - Sorry, we can't find the page you are looking for`. You must be [logged in](https://huggingface.co/docs/huggingface_hub/quick-start#login) to load a model from a private repository.
+Private repositories are only visible to you, and other users won't be able to clone the repository and your repository won't appear in search results. Even if a user has the URL to your private repository, they'll receive a `404 - Sorry, we can't find the page you are looking for.`
+
+To load a model, scheduler, or pipeline from private or gated repositories, set `use_auth_token=True`:
+
+```py
+model = ControlNetModel.from_pretrained("your-namespace/my-controlnet-model-private", use_auth_token=True)
+```

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -123,29 +123,20 @@ def save_model_card(
         """
 
     trigger_str = f"You should use {instance_prompt} to trigger the image generation."
-    diffusers_imports_pivotal = ""
-    diffusers_example_pivotal = ""
     if train_text_encoder_ti:
         trigger_str = (
             "To trigger image generation of trained concept(or concepts) replace each concept identifier "
             "in you prompt with the new inserted tokens:\n"
         )
-        diffusers_imports_pivotal = """from huggingface_hub import hf_hub_download
-from safetensors.torch import load_file
-        """
-        diffusers_example_pivotal = f"""embedding_path = hf_hub_download(repo_id='{repo_id}', filename="embeddings.safetensors", repo_type="model")
-state_dict = load_file(embedding_path)
-pipeline.load_textual_inversion(state_dict["clip_l"], token=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
-pipeline.load_textual_inversion(state_dict["clip_g"], token=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
-        """
         if token_abstraction_dict:
             for key, value in token_abstraction_dict.items():
                 tokens = "".join(value)
                 trigger_str += f"""
-to trigger concept `{key}` → use `{tokens}` in your prompt \n
+to trigger concept `{key}->` use `{tokens}` in your prompt \n
 """
 
-    yaml = f"""---
+    yaml = f"""
+---
 tags:
 - stable-diffusion-xl
 - stable-diffusion-xl-diffusers
@@ -157,10 +148,8 @@ tags:
 base_model: {base_model}
 instance_prompt: {instance_prompt}
 license: openrail++
-widget:
-    - text: '{validation_prompt if validation_prompt else instance_prompt}'
 ---
-"""
+    """
 
     model_card = f"""
 # SDXL LoRA DreamBooth - {repo_id}
@@ -171,43 +160,24 @@ widget:
 
 ### These are {repo_id} LoRA adaption weights for {base_model}.
 
-## Trigger words
+The weights were trained  using [DreamBooth](https://dreambooth.github.io/).
 
-{trigger_str}
-
-## Use it with the [🧨 diffusers library](https://github.com/huggingface/diffusers)
-
-```py
-from diffusers import AutoPipelineForText2Image
-import torch
-{diffusers_imports_pivotal}
-pipeline = AutoPipelineForText2Image.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', torch_dtype=torch.float16).to('cuda')
-pipeline.load_lora_weights('{repo_id}', weight_name='pytorch_lora_weights.safetensors')
-{diffusers_example_pivotal}
-image = pipeline('{validation_prompt if validation_prompt else instance_prompt}').images[0]
-```
-
-For more details, including weighting, merging and fusing LoRAs, check the [documentation on loading LoRAs in diffusers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/loading_adapters)
-
-## Download model
-
-### Use it with UIs such as AUTOMATIC1111, Comfy UI, SD.Next, Invoke
-
-- Download the LoRA *.safetensors [here](/{repo_id}/blob/main/pytorch_lora_weights.safetensors). Rename it and place it on your Lora folder.
-- Download the text embeddings *.safetensors [here](/{repo_id}/blob/main/embeddings.safetensors). Rename it and place it on it on your embeddings folder.
-
-All [Files & versions](/{repo_id}/tree/main).
-
-## Details
-
-The weights were trained using [🧨 diffusers Advanced Dreambooth Training Script](https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py).
-
-LoRA for the text encoder was enabled. {train_text_encoder}.
+LoRA for the text encoder was enabled: {train_text_encoder}.
 
 Pivotal tuning was enabled: {train_text_encoder_ti}.
 
 Special VAE used for training: {vae_path}.
 
+## Trigger words
+
+{trigger_str}
+
+## Download model
+
+Weights for this model are available in Safetensors format.
+
+[Download]({repo_id}/tree/main) them in the Files & versions tab.
+
 """
     with open(os.path.join(repo_folder, "README.md"), "w") as f:
         f.write(yaml + model_card)
@@ -673,12 +643,6 @@ def parse_args(input_args=None):
         default=4,
         help=("The dimension of the LoRA update matrices."),
     )
-    parser.add_argument(
-        "--cache_latents",
-        action="store_true",
-        default=False,
-        help="Cache the VAE latents",
-    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -827,12 +791,6 @@ class DreamBoothDataset(Dataset):
         instance_data_root,
         instance_prompt,
         class_prompt,
-        dataset_name,
-        dataset_config_name,
-        cache_dir,
-        image_column,
-        caption_column,
-        train_text_encoder_ti,
         class_data_root=None,
         class_num=None,
         token_abstraction_dict=None,  # token mapping for textual inversion
@@ -847,10 +805,10 @@ class DreamBoothDataset(Dataset):
         self.custom_instance_prompts = None
         self.class_prompt = class_prompt
         self.token_abstraction_dict = token_abstraction_dict
-        self.train_text_encoder_ti = train_text_encoder_ti
+
         # if --dataset_name is provided or a metadata jsonl file is provided in the local --instance_data directory,
         # we load the training data using load_dataset
-        if dataset_name is not None:
+        if args.dataset_name is not None:
             try:
                 from datasets import load_dataset
             except ImportError:
@@ -863,25 +821,26 @@ class DreamBoothDataset(Dataset):
             # See more about loading custom images at
             # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
             dataset = load_dataset(
-                dataset_name,
-                dataset_config_name,
-                cache_dir=cache_dir,
+                args.dataset_name,
+                args.dataset_config_name,
+                cache_dir=args.cache_dir,
             )
             # Preprocessing the datasets.
             column_names = dataset["train"].column_names
 
             # 6. Get the column names for input/target.
-            if image_column is None:
+            if args.image_column is None:
                 image_column = column_names[0]
                 logger.info(f"image column defaulting to {image_column}")
             else:
+                image_column = args.image_column
                 if image_column not in column_names:
                     raise ValueError(
-                        f"`--image_column` value '{image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+                        f"`--image_column` value '{args.image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
                     )
             instance_images = dataset["train"][image_column]
 
-            if caption_column is None:
+            if args.caption_column is None:
                 logger.info(
                     "No caption column provided, defaulting to instance_prompt for all images. If your dataset "
                     "contains captions/prompts for the images, make sure to specify the "
@@ -889,11 +848,11 @@ class DreamBoothDataset(Dataset):
                 )
                 self.custom_instance_prompts = None
             else:
-                if caption_column not in column_names:
+                if args.caption_column not in column_names:
                     raise ValueError(
-                        f"`--caption_column` value '{caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+                        f"`--caption_column` value '{args.caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
                     )
-                custom_instance_prompts = dataset["train"][caption_column]
+                custom_instance_prompts = dataset["train"][args.caption_column]
                 # create final list of captions according to --repeats
                 self.custom_instance_prompts = []
                 for caption in custom_instance_prompts:
@@ -948,7 +907,7 @@ class DreamBoothDataset(Dataset):
         if self.custom_instance_prompts:
             caption = self.custom_instance_prompts[index % self.num_instance_images]
             if caption:
-                if self.train_text_encoder_ti:
+                if args.train_text_encoder_ti:
                     # replace instances of --token_abstraction in caption with the new tokens: "<si><si+1>" etc.
                     for token_abs, token_replacement in self.token_abstraction_dict.items():
                         caption = caption.replace(token_abs, "".join(token_replacement))
@@ -1134,10 +1093,10 @@ def main(args):
         if args.output_dir is not None:
             os.makedirs(args.output_dir, exist_ok=True)
 
-        model_id = args.hub_model_id or Path(args.output_dir).name
-        repo_id = None
         if args.push_to_hub:
-            repo_id = create_repo(repo_id=model_id, exist_ok=True, token=args.hub_token).repo_id
+            repo_id = create_repo(
+                repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
+            ).repo_id
 
     # Load the tokenizers
     tokenizer_one = AutoTokenizer.from_pretrained(
@@ -1182,7 +1141,6 @@ def main(args):
         revision=args.revision,
         variant=args.variant,
     )
-    vae_scaling_factor = vae.config.scaling_factor
     unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
     )
@@ -1506,12 +1464,6 @@ def main(args):
         instance_data_root=args.instance_data_dir,
         instance_prompt=args.instance_prompt,
         class_prompt=args.class_prompt,
-        dataset_name=args.dataset_name,
-        dataset_config_name=args.dataset_config_name,
-        cache_dir=args.cache_dir,
-        image_column=args.image_column,
-        train_text_encoder_ti=args.train_text_encoder_ti,
-        caption_column=args.caption_column,
         class_data_root=args.class_data_dir if args.with_prior_preservation else None,
         token_abstraction_dict=token_abstraction_dict if args.train_text_encoder_ti else None,
         class_num=args.num_class_images,
@@ -1613,20 +1565,6 @@ def main(args):
             args.validation_prompt = args.validation_prompt.replace(token_abs, "".join(token_replacement))
     print("validation prompt:", args.validation_prompt)
 
-    if args.cache_latents:
-        latents_cache = []
-        for batch in tqdm(train_dataloader, desc="Caching latents"):
-            with torch.no_grad():
-                batch["pixel_values"] = batch["pixel_values"].to(
-                    accelerator.device, non_blocking=True, dtype=torch.float32
-                )
-                latents_cache.append(vae.encode(batch["pixel_values"]).latent_dist)
-
-        if args.validation_prompt is None:
-            del vae
-            if torch.cuda.is_available():
-                torch.cuda.empty_cache()
-
     # Scheduler and math around the number of training steps.
     overrode_max_train_steps = False
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
@@ -1742,7 +1680,9 @@ def main(args):
         unet.train()
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
+                pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                 prompts = batch["prompts"]
+                # print(prompts)
                 # encode batch prompts when custom prompts are provided for each image -
                 if train_dataset.custom_instance_prompts:
                     if freeze_text_encoder:
@@ -1754,13 +1694,9 @@ def main(args):
                         tokens_one = tokenize_prompt(tokenizer_one, prompts, add_special_tokens)
                         tokens_two = tokenize_prompt(tokenizer_two, prompts, add_special_tokens)
 
-                if args.cache_latents:
-                    model_input = latents_cache[step].sample()
-                else:
-                    pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
-                    model_input = vae.encode(pixel_values).latent_dist.sample()
-
-                model_input = model_input * vae_scaling_factor
+                # Convert images to latent space
+                model_input = vae.encode(pixel_values).latent_dist.sample()
+                model_input = model_input * vae.config.scaling_factor
                 if args.pretrained_vae_model_name_or_path is None:
                     model_input = model_input.to(weight_dtype)
 
@@ -2068,23 +2004,23 @@ def main(args):
                         }
                     )
 
-        if args.train_text_encoder_ti:
-            embedding_handler.save_embeddings(
-                f"{args.output_dir}/embeddings.safetensors",
-            )
-        save_model_card(
-            model_id if not args.push_to_hub else repo_id,
-            images=images,
-            base_model=args.pretrained_model_name_or_path,
-            train_text_encoder=args.train_text_encoder,
-            train_text_encoder_ti=args.train_text_encoder_ti,
-            token_abstraction_dict=train_dataset.token_abstraction_dict,
-            instance_prompt=args.instance_prompt,
-            validation_prompt=args.validation_prompt,
-            repo_folder=args.output_dir,
-            vae_path=args.pretrained_vae_model_name_or_path,
-        )
         if args.push_to_hub:
+            if args.train_text_encoder_ti:
+                embedding_handler.save_embeddings(
+                    f"{args.output_dir}/embeddings.safetensors",
+                )
+            save_model_card(
+                repo_id,
+                images=images,
+                base_model=args.pretrained_model_name_or_path,
+                train_text_encoder=args.train_text_encoder,
+                train_text_encoder_ti=args.train_text_encoder_ti,
+                token_abstraction_dict=train_dataset.token_abstraction_dict,
+                instance_prompt=args.instance_prompt,
+                validation_prompt=args.validation_prompt,
+                repo_folder=args.output_dir,
+                vae_path=args.pretrained_vae_model_name_or_path,
+            )
             upload_folder(
                 repo_id=repo_id,
                 folder_path=args.output_dir,

examples/community/README.md

@@ -50,7 +50,6 @@ prompt-to-prompt | change parts of a prompt and retain image structure (see [pap
 |   Latent Consistency Interpolation Pipeline                                                                                                    | Interpolate the latent space of Latent Consistency Models with multiple prompts                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Interpolation Pipeline](#latent-consistency-interpolation-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1pK3NrLWJSiJsBynLns1K1-IDTW9zbPvl?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
 |   Regional Prompting Pipeline                                                                                               | Assign multiple prompts for different regions                                                                                                                                                                                                                                                                                                                                                    |  [Regional Prompting Pipeline](#regional-prompting-pipeline) | - | [hako-mikan](https://github.com/hako-mikan) |
 | LDM3D-sr (LDM3D upscaler)                                                                                                             | Upscale low resolution RGB and depth inputs to high resolution                                                                                                                                                                                                                                                                                                                                                                                                                              | [StableDiffusionUpscaleLDM3D Pipeline](https://github.com/estelleafl/diffusers/tree/ldm3d_upscaler_community/examples/community#stablediffusionupscaleldm3d-pipeline)                                                                             | -                                                                                                                                                                                                             |                                                        [Estelle Aflalo](https://github.com/estelleafl) |
-| AnimateDiff ControlNet Pipeline                                                                                                    | Combines AnimateDiff with precise motion control using ControlNets                                                                                                                                                                                                                                                                                                                                                                                                                                    | [AnimateDiff ControlNet Pipeline](#animatediff-controlnet-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SKboYeGjEQmQPWoFC0aLYpBlYdHXkvAu?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) and [Edoardo Botta](https://github.com/EdoardoBotta) |
 |   DemoFusion Pipeline                                                                                                    | Implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [DemoFusion Pipeline](#DemoFusion)      | - |              [Ruoyi Du](https://github.com/RuoyiDu) |
 
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
@@ -513,6 +512,7 @@ device = torch.device('cpu' if not has_cuda else 'cuda')
 pipe = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
         safety_checker=None,
+    use_auth_token=True,
     custom_pipeline="imagic_stable_diffusion",
     scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
 ).to(device)
@@ -552,6 +552,7 @@ device = th.device('cpu' if not has_cuda else 'cuda')
 
 pipe = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
+    use_auth_token=True,
     custom_pipeline="seed_resize_stable_diffusion"
 ).to(device)
 
@@ -587,6 +588,7 @@ generator = th.Generator("cuda").manual_seed(0)
 
 pipe = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
+    use_auth_token=True,
     custom_pipeline="/home/mark/open_source/diffusers/examples/community/"
 ).to(device)
 
@@ -605,6 +607,7 @@ image.save('./seed_resize/seed_resize_{w}_{h}_image.png'.format(w=width, h=heigh
 
 pipe_compare = DiffusionPipeline.from_pretrained(
     "CompVis/stable-diffusion-v1-4",
+    use_auth_token=True,
     custom_pipeline="/home/mark/open_source/diffusers/examples/community/"
 ).to(device)
 
@@ -2840,70 +2843,6 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
     * Reconstructed image:
     * ![dps_generated_image](https://github.com/tongdaxu/Images/assets/22267548/b74f084d-93f4-4845-83d8-44c0fa758a5f)
 
-### AnimateDiff ControlNet Pipeline
-
-This pipeline combines AnimateDiff and ControlNet. Enjoy precise motion control for your videos! Refer to [this](https://github.com/huggingface/diffusers/issues/5866) issue for more details.
-
-```py
-import torch
-from diffusers import AutoencoderKL, ControlNetModel, MotionAdapter
-from diffusers.pipelines import DiffusionPipeline
-from diffusers.schedulers import DPMSolverMultistepScheduler
-from PIL import Image
-
-motion_id = "guoyww/animatediff-motion-adapter-v1-5-2"
-adapter = MotionAdapter.from_pretrained(motion_id)
-controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_openpose", torch_dtype=torch.float16)
-vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16)
-
-model_id = "SG161222/Realistic_Vision_V5.1_noVAE"
-pipe = DiffusionPipeline.from_pretrained(
-    model_id,
-    motion_adapter=adapter,
-    controlnet=controlnet,
-    vae=vae,
-    custom_pipeline="pipeline_animatediff_controlnet",
-).to(device="cuda", dtype=torch.float16)
-pipe.scheduler = DPMSolverMultistepScheduler.from_pretrained(
-    model_id, subfolder="scheduler", clip_sample=False, timestep_spacing="linspace", steps_offset=1
-)
-pipe.enable_vae_slicing()
-
-conditioning_frames = []
-for i in range(1, 16 + 1):
-    conditioning_frames.append(Image.open(f"frame_{i}.png"))
-
-prompt = "astronaut in space, dancing"
-negative_prompt = "bad quality, worst quality, jpeg artifacts, ugly"
-result = pipe(
-    prompt=prompt,
-    negative_prompt=negative_prompt,
-    width=512,
-    height=768,
-    conditioning_frames=conditioning_frames,
-    num_inference_steps=12,
-).frames[0]
-
-from diffusers.utils import export_to_gif
-export_to_gif(result.frames[0], "result.gif")
-```
-
-<table>
-  <tr><td colspan="2" align=center><b>Conditioning Frames</b></td></tr>
-  <tr align=center>
-    <td align=center><img src="https://user-images.githubusercontent.com/7365912/265043418-23291941-864d-495a-8ba8-d02e05756396.gif" alt="input-frames"></td>
-  </tr>
-  <tr><td colspan="2" align=center><b>AnimateDiff model: SG161222/Realistic_Vision_V5.1_noVAE</b></td></tr>
-  <tr>
-    <td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/baf301e2-d03c-4129-bd84-203a1de2b2be" alt="gif-1"></td>
-    <td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/9f923475-ecaf-452b-92c8-4e42171182d8" alt="gif-2"></td>
-  </tr>
-  <tr><td colspan="2" align=center><b>AnimateDiff model: CardosAnime</b></td></tr>
-  <tr>
-    <td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/b2c41028-38a0-45d6-86ed-fec7446b87f7" alt="gif-1"></td>
-    <td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/eb7d2952-72e4-44fa-b664-077c79b4fc70" alt="gif-2"></td>
-  </tr>
-</table>
 ### DemoFusion
 This pipeline is the official implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973).
 The original repo can be found at [repo](https://github.com/PRIS-CV/DemoFusion).
@@ -2931,14 +2870,10 @@ The original repo can be found at [repo](https://github.com/PRIS-CV/DemoFusion).
 - `show_image` (`bool`, defaults to False):
   Determine whether to show intermediate results during generation.
 ```
-from diffusers import DiffusionPipeline
+from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
 
-pipe = DiffusionPipeline.from_pretrained(
-    "stabilityai/stable-diffusion-xl-base-1.0",
-    custom_pipeline="pipeline_demofusion_sdxl",
-    custom_revision="main",
-    torch_dtype=torch.float16,
-)
+model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
 pipe = pipe.to("cuda")
 
 prompt = "Envision a portrait of an elderly woman, her face a canvas of time, framed by a headscarf with muted tones of rust and cream. Her eyes, blue like faded denim. Her attire, simple yet dignified."

examples/community/checkpoint_merger.py

@@ -5,11 +5,10 @@ from typing import Dict, List, Union
 import safetensors.torch
 import torch
 from huggingface_hub import snapshot_download
-from huggingface_hub.utils import validate_hf_hub_args
 
 from diffusers import DiffusionPipeline, __version__
 from diffusers.schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
-from diffusers.utils import CONFIG_NAME, ONNX_WEIGHTS_NAME, WEIGHTS_NAME
+from diffusers.utils import CONFIG_NAME, DIFFUSERS_CACHE, ONNX_WEIGHTS_NAME, WEIGHTS_NAME
 
 
 class CheckpointMergerPipeline(DiffusionPipeline):
@@ -58,7 +57,6 @@ class CheckpointMergerPipeline(DiffusionPipeline):
         return (temp_dict, meta_keys)
 
     @torch.no_grad()
-    @validate_hf_hub_args
     def merge(self, pretrained_model_name_or_path_list: List[Union[str, os.PathLike]], **kwargs):
         """
         Returns a new pipeline object of the class 'DiffusionPipeline' with the merged checkpoints(weights) of the models passed
@@ -71,7 +69,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
             **kwargs:
                 Supports all the default DiffusionPipeline.get_config_dict kwargs viz..
 
-                cache_dir, resume_download, force_download, proxies, local_files_only, token, revision, torch_dtype, device_map.
+                cache_dir, resume_download, force_download, proxies, local_files_only, use_auth_token, revision, torch_dtype, device_map.
 
                 alpha - The interpolation parameter. Ranges from 0 to 1.  It affects the ratio in which the checkpoints are merged. A 0.8 alpha
                     would mean that the first model checkpoints would affect the final result far less than an alpha of 0.2
@@ -83,12 +81,12 @@ class CheckpointMergerPipeline(DiffusionPipeline):
 
         """
         # Default kwargs from DiffusionPipeline
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
         device_map = kwargs.pop("device_map", None)
@@ -125,7 +123,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
                 force_download=force_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
             )
             config_dicts.append(config_dict)
@@ -161,7 +159,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     revision=revision,
                     allow_patterns=allow_patterns,
                     user_agent=user_agent,

examples/community/pipeline_demofusion_sdxl.py

@@ -36,9 +36,7 @@ from diffusers.utils.torch_utils import randn_tensor
 
 
 if is_invisible_watermark_available():
-    from diffusers.pipelines.stable_diffusion_xl.watermark import (
-        StableDiffusionXLWatermarker,
-    )
+    from .watermark import StableDiffusionXLWatermarker
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_diffusion_tensorrt_img2img.py

@@ -28,7 +28,6 @@ import PIL.Image
 import tensorrt as trt
 import torch
 from huggingface_hub import snapshot_download
-from huggingface_hub.utils import validate_hf_hub_args
 from onnx import shape_inference
 from polygraphy import cuda
 from polygraphy.backend.common import bytes_from_path
@@ -51,7 +50,7 @@ from diffusers.pipelines.stable_diffusion import (
     StableDiffusionSafetyChecker,
 )
 from diffusers.schedulers import DDIMScheduler
-from diffusers.utils import logging
+from diffusers.utils import DIFFUSERS_CACHE, logging
 
 
 """
@@ -779,13 +778,12 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             self.models["vae_encoder"] = make_VAEEncoder(self.vae, **models_args)
 
     @classmethod
-    @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
 
         cls.cached_folder = (
@@ -797,7 +795,7 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
             )
         )

examples/community/stable_diffusion_tensorrt_inpaint.py

@@ -28,7 +28,6 @@ import PIL.Image
 import tensorrt as trt
 import torch
 from huggingface_hub import snapshot_download
-from huggingface_hub.utils import validate_hf_hub_args
 from onnx import shape_inference
 from polygraphy import cuda
 from polygraphy.backend.common import bytes_from_path
@@ -52,7 +51,7 @@ from diffusers.pipelines.stable_diffusion import (
 )
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint import prepare_mask_and_masked_image
 from diffusers.schedulers import DDIMScheduler
-from diffusers.utils import logging
+from diffusers.utils import DIFFUSERS_CACHE, logging
 
 
 """
@@ -780,13 +779,12 @@ class TensorRTStableDiffusionInpaintPipeline(StableDiffusionInpaintPipeline):
             self.models["vae_encoder"] = make_VAEEncoder(self.vae, **models_args)
 
     @classmethod
-    @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
 
         cls.cached_folder = (
@@ -798,7 +796,7 @@ class TensorRTStableDiffusionInpaintPipeline(StableDiffusionInpaintPipeline):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
             )
         )

examples/community/stable_diffusion_tensorrt_txt2img.py

@@ -27,7 +27,6 @@ import onnx_graphsurgeon as gs
 import tensorrt as trt
 import torch
 from huggingface_hub import snapshot_download
-from huggingface_hub.utils import validate_hf_hub_args
 from onnx import shape_inference
 from polygraphy import cuda
 from polygraphy.backend.common import bytes_from_path
@@ -50,7 +49,7 @@ from diffusers.pipelines.stable_diffusion import (
     StableDiffusionSafetyChecker,
 )
 from diffusers.schedulers import DDIMScheduler
-from diffusers.utils import logging
+from diffusers.utils import DIFFUSERS_CACHE, logging
 
 
 """
@@ -692,13 +691,12 @@ class TensorRTStableDiffusionPipeline(StableDiffusionPipeline):
             self.models["vae"] = make_VAE(self.vae, **models_args)
 
     @classmethod
-    @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
 
         cls.cached_folder = (
@@ -710,7 +708,7 @@ class TensorRTStableDiffusionPipeline(StableDiffusionPipeline):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
             )
         )

examples/consistency_distillation/README.md

@@ -1,6 +1,6 @@
 # Latent Consistency Distillation Example:
 
-[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is a method to distill a latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use latent consistency distillation to distill stable-diffusion-v1.5 for inference with few timesteps.
+[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is method to distill latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use the latent consistency distillation to distill stable-diffusion-v1.5 for less timestep inference.
 
 ## Full model distillation
 
@@ -24,7 +24,7 @@ Then cd in the example folder and run
 pip install -r requirements.txt
 ```
 
-And initialize an [🤗 Accelerate](https://github.com/huggingface/accelerate/) environment with:
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
 
 ```bash
 accelerate config
@@ -46,16 +46,12 @@ write_basic_config()
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
 
 
-#### Example
-
-The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example, and for illustrative purposes only. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/). You may also need to search the hyperparameter space according to the dataset you use.
+#### Example with LAION-A6+ dataset
 
 ```bash
-export MODEL_NAME="runwayml/stable-diffusion-v1-5"
-export OUTPUT_DIR="path/to/saved/model"
-
-accelerate launch train_lcm_distill_sd_wds.py \
-    --pretrained_teacher_model=$MODEL_NAME \
+runwayml/stable-diffusion-v1-5
+PROGRAM="train_lcm_distill_sd_wds.py \
+    --pretrained_teacher_model=$MODEL_DIR \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
     --resolution=512 \
@@ -63,7 +59,7 @@ accelerate launch train_lcm_distill_sd_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
+    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \
@@ -73,23 +69,19 @@ accelerate launch train_lcm_distill_sd_wds.py \
     --resume_from_checkpoint=latest \
     --report_to=wandb \
     --seed=453645634 \
-    --push_to_hub
+    --push_to_hub \
 ```
 
 ## LCM-LoRA
 
 Instead of fine-tuning the full model, we can also just train a LoRA that can be injected into any SDXL model.
 
-### Example
-
-The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/).
-
+### Example with LAION-A6+ dataset
+    
 ```bash
-export MODEL_NAME="runwayml/stable-diffusion-v1-5"
-export OUTPUT_DIR="path/to/saved/model"
-
-accelerate launch train_lcm_distill_lora_sd_wds.py \
-    --pretrained_teacher_model=$MODEL_NAME \
+runwayml/stable-diffusion-v1-5
+PROGRAM="train_lcm_distill_lora_sd_wds.py \
+    --pretrained_teacher_model=$MODEL_DIR \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
     --resolution=512 \
@@ -98,7 +90,7 @@ accelerate launch train_lcm_distill_lora_sd_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
+    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \

examples/consistency_distillation/README_sdxl.md

@@ -1,6 +1,6 @@
 # Latent Consistency Distillation Example:
 
-[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is a method to distill a latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use latent consistency distillation to distill SDXL for inference with few timesteps.
+[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is method to distill latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use the latent consistency distillation to distill SDXL for less timestep inference.
 
 ## Full model distillation
 
@@ -24,7 +24,7 @@ Then cd in the example folder and run
 pip install -r requirements.txt
 ```
 
-And initialize an [🤗 Accelerate](https://github.com/huggingface/accelerate/) environment with:
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
 
 ```bash
 accelerate config
@@ -46,16 +46,12 @@ write_basic_config()
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
 
 
-#### Example
-
-The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example, and for illustrative purposes only. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/). You may also need to search the hyperparameter space according to the dataset you use.
+#### Example with LAION-A6+ dataset
 
 ```bash
-export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
-export OUTPUT_DIR="path/to/saved/model"
-
-accelerate launch train_lcm_distill_sdxl_wds.py \
-    --pretrained_teacher_model=$MODEL_NAME \
+export MODEL_DIR="stabilityai/stable-diffusion-xl-base-1.0"
+PROGRAM="train_lcm_distill_sdxl_wds.py \
+    --pretrained_teacher_model=$MODEL_DIR \
     --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
@@ -64,7 +60,7 @@ accelerate launch train_lcm_distill_sdxl_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
+    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \
@@ -81,15 +77,11 @@ accelerate launch train_lcm_distill_sdxl_wds.py \
 
 Instead of fine-tuning the full model, we can also just train a LoRA that can be injected into any SDXL model.
 
-### Example
-
-The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/).
-
+### Example with LAION-A6+ dataset
+    
 ```bash
-export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
-export OUTPUT_DIR="path/to/saved/model"
-
-accelerate launch train_lcm_distill_lora_sdxl_wds.py \
+export MODEL_DIR="stabilityai/stable-diffusion-xl-base-1.0"
+PROGRAM="train_lcm_distill_lora_sdxl_wds.py \
     --pretrained_teacher_model=$MODEL_DIR \
     --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
     --output_dir=$OUTPUT_DIR \
@@ -100,7 +92,7 @@ accelerate launch train_lcm_distill_lora_sdxl_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
+    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \

examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py

@@ -423,7 +423,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
     )
     model_class = text_encoder_config.architectures[0]
 
@@ -1123,7 +1123,7 @@ def main(args):
     for epoch in range(first_epoch, args.num_train_epochs):
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
-                image, text = batch
+                image, text, _, _ = batch
 
                 image = image.to(accelerator.device, non_blocking=True)
                 encoded_text = compute_embeddings_fn(text)

examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

@@ -68,11 +68,6 @@ from diffusers.utils.import_utils import is_xformers_available
 
 MAX_SEQ_LENGTH = 77
 
-# Adjust for your dataset
-WDS_JSON_WIDTH = "width"  # original_width for LAION
-WDS_JSON_HEIGHT = "height"  # original_height for LAION
-MIN_SIZE = 700  # ~960 for LAION, ideal: 1024 if the dataset contains large images
-
 if is_wandb_available():
     import wandb
 
@@ -151,10 +146,10 @@ class WebdatasetFilter:
         try:
             if "json" in x:
                 x_json = json.loads(x["json"])
-                filter_size = (x_json.get(WDS_JSON_WIDTH, 0.0) or 0.0) >= self.min_size and x_json.get(
-                    WDS_JSON_HEIGHT, 0
+                filter_size = (x_json.get("original_width", 0.0) or 0.0) >= self.min_size and x_json.get(
+                    "original_height", 0
                 ) >= self.min_size
-                filter_watermark = (x_json.get("pwatermark", 0.0) or 0.0) <= self.max_pwatermark
+                filter_watermark = (x_json.get("pwatermark", 1.0) or 1.0) <= self.max_pwatermark
                 return filter_size and filter_watermark
             else:
                 return False
@@ -185,7 +180,7 @@ class Text2ImageDataset:
             if use_fix_crop_and_size:
                 return (resolution, resolution)
             else:
-                return (int(json.get(WDS_JSON_WIDTH, 0.0)), int(json.get(WDS_JSON_HEIGHT, 0.0)))
+                return (int(json.get("original_width", 0.0)), int(json.get("original_height", 0.0)))
 
         def transform(example):
             # resize image
@@ -217,7 +212,7 @@ class Text2ImageDataset:
         pipeline = [
             wds.ResampledShards(train_shards_path_or_url),
             tarfile_to_samples_nothrow,
-            wds.select(WebdatasetFilter(min_size=MIN_SIZE)),
+            wds.select(WebdatasetFilter(min_size=960)),
             wds.shuffle(shuffle_buffer_size),
             *processing_pipeline,
             wds.batched(per_gpu_batch_size, partial=False, collation_fn=default_collate),
@@ -397,7 +392,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
     )
     model_class = text_encoder_config.architectures[0]
 

examples/consistency_distillation/train_lcm_distill_sd_wds.py

@@ -400,7 +400,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
     )
     model_class = text_encoder_config.architectures[0]
 
@@ -1106,7 +1106,7 @@ def main(args):
     for epoch in range(first_epoch, args.num_train_epochs):
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
-                image, text = batch
+                image, text, _, _ = batch
 
                 image = image.to(accelerator.device, non_blocking=True)
                 encoded_text = compute_embeddings_fn(text)

examples/consistency_distillation/train_lcm_distill_sdxl_wds.py

@@ -67,11 +67,6 @@ from diffusers.utils.import_utils import is_xformers_available
 
 MAX_SEQ_LENGTH = 77
 
-# Adjust for your dataset
-WDS_JSON_WIDTH = "width"  # original_width for LAION
-WDS_JSON_HEIGHT = "height"  # original_height for LAION
-MIN_SIZE = 700  # ~960 for LAION, ideal: 1024 if the dataset contains large images
-
 if is_wandb_available():
     import wandb
 
@@ -133,10 +128,10 @@ class WebdatasetFilter:
         try:
             if "json" in x:
                 x_json = json.loads(x["json"])
-                filter_size = (x_json.get(WDS_JSON_WIDTH, 0.0) or 0.0) >= self.min_size and x_json.get(
-                    WDS_JSON_HEIGHT, 0
+                filter_size = (x_json.get("original_width", 0.0) or 0.0) >= self.min_size and x_json.get(
+                    "original_height", 0
                 ) >= self.min_size
-                filter_watermark = (x_json.get("pwatermark", 0.0) or 0.0) <= self.max_pwatermark
+                filter_watermark = (x_json.get("pwatermark", 1.0) or 1.0) <= self.max_pwatermark
                 return filter_size and filter_watermark
             else:
                 return False
@@ -167,7 +162,7 @@ class Text2ImageDataset:
             if use_fix_crop_and_size:
                 return (resolution, resolution)
             else:
-                return (int(json.get(WDS_JSON_WIDTH, 0.0)), int(json.get(WDS_JSON_HEIGHT, 0.0)))
+                return (int(json.get("original_width", 0.0)), int(json.get("original_height", 0.0)))
 
         def transform(example):
             # resize image
@@ -199,7 +194,7 @@ class Text2ImageDataset:
         pipeline = [
             wds.ResampledShards(train_shards_path_or_url),
             tarfile_to_samples_nothrow,
-            wds.select(WebdatasetFilter(min_size=MIN_SIZE)),
+            wds.select(WebdatasetFilter(min_size=960)),
             wds.shuffle(shuffle_buffer_size),
             *processing_pipeline,
             wds.batched(per_gpu_batch_size, partial=False, collation_fn=default_collate),
@@ -419,7 +414,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
     )
     model_class = text_encoder_config.architectures[0]
 

examples/dreambooth/README.md

@@ -44,7 +44,6 @@ write_basic_config()
 ```
 
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups. 
-Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
 
 ### Dog toy example
 

examples/dreambooth/README_sdxl.md

@@ -47,7 +47,6 @@ write_basic_config()
 ```
 
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups. 
-Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
 
 ### Dog toy example
 

examples/dreambooth/requirements.txt

@@ -4,4 +4,3 @@ transformers>=4.25.1
 ftfy
 tensorboard
 Jinja2
-peft==0.7.0

examples/dreambooth/requirements_sdxl.txt

@@ -4,4 +4,3 @@ transformers>=4.25.1
 ftfy
 tensorboard
 Jinja2
-peft==0.7.0

examples/dreambooth/train_dreambooth_lora.py

@@ -16,6 +16,7 @@
 import argparse
 import copy
 import gc
+import itertools
 import logging
 import math
 import os
@@ -34,8 +35,6 @@ from accelerate.utils import ProjectConfiguration, set_seed
 from huggingface_hub import create_repo, upload_folder
 from huggingface_hub.utils import insecure_hashlib
 from packaging import version
-from peft import LoraConfig
-from peft.utils import get_peft_model_state_dict
 from PIL import Image
 from PIL.ImageOps import exif_transpose
 from torch.utils.data import Dataset
@@ -53,7 +52,14 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.loaders import LoraLoaderMixin
+from diffusers.models.attention_processor import (
+    AttnAddedKVProcessor,
+    AttnAddedKVProcessor2_0,
+    SlicedAttnAddedKVProcessor,
+)
+from diffusers.models.lora import LoRALinearLayer
 from diffusers.optimization import get_scheduler
+from diffusers.training_utils import unet_lora_state_dict
 from diffusers.utils import check_min_version, is_wandb_available
 from diffusers.utils.import_utils import is_xformers_available
 
@@ -858,19 +864,79 @@ def main(args):
             text_encoder.gradient_checkpointing_enable()
 
     # now we will add new LoRA weights to the attention layers
-    unet_lora_config = LoraConfig(
-        r=args.rank,
-        init_lora_weights="gaussian",
-        target_modules=["to_k", "to_q", "to_v", "to_out.0", "add_k_proj", "add_v_proj"],
-    )
-    unet.add_adapter(unet_lora_config)
+    # It's important to realize here how many attention weights will be added and of which sizes
+    # The sizes of the attention layers consist only of two different variables:
+    # 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.
+    # 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.
 
-    # The text encoder comes from 🤗 transformers, we will also attach adapters to it.
-    if args.train_text_encoder:
-        text_lora_config = LoraConfig(
-            r=args.rank, init_lora_weights="gaussian", target_modules=["q_proj", "k_proj", "v_proj", "out_proj"]
+    # Let's first see how many attention processors we will have to set.
+    # For Stable Diffusion, it should be equal to:
+    # - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12
+    # - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2
+    # - up blocks (2x attention layers) * (3x transformer layers) * (3x up blocks) = 18
+    # => 32 layers
+
+    # Set correct lora layers
+    unet_lora_parameters = []
+    for attn_processor_name, attn_processor in unet.attn_processors.items():
+        # Parse the attention module.
+        attn_module = unet
+        for n in attn_processor_name.split(".")[:-1]:
+            attn_module = getattr(attn_module, n)
+
+        # Set the `lora_layer` attribute of the attention-related matrices.
+        attn_module.to_q.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
+            )
         )
-        text_encoder.add_adapter(text_lora_config)
+        attn_module.to_k.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_v.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_out[0].set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_out[0].in_features,
+                out_features=attn_module.to_out[0].out_features,
+                rank=args.rank,
+            )
+        )
+
+        # Accumulate the LoRA params to optimize.
+        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
+
+        if isinstance(attn_processor, (AttnAddedKVProcessor, SlicedAttnAddedKVProcessor, AttnAddedKVProcessor2_0)):
+            attn_module.add_k_proj.set_lora_layer(
+                LoRALinearLayer(
+                    in_features=attn_module.add_k_proj.in_features,
+                    out_features=attn_module.add_k_proj.out_features,
+                    rank=args.rank,
+                )
+            )
+            attn_module.add_v_proj.set_lora_layer(
+                LoRALinearLayer(
+                    in_features=attn_module.add_v_proj.in_features,
+                    out_features=attn_module.add_v_proj.out_features,
+                    rank=args.rank,
+                )
+            )
+            unet_lora_parameters.extend(attn_module.add_k_proj.lora_layer.parameters())
+            unet_lora_parameters.extend(attn_module.add_v_proj.lora_layer.parameters())
+
+    # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
+    # So, instead, we monkey-patch the forward calls of its attention-blocks.
+    if args.train_text_encoder:
+        # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
+        text_lora_parameters = LoraLoaderMixin._modify_text_encoder(text_encoder, dtype=torch.float32, rank=args.rank)
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
@@ -882,9 +948,9 @@ def main(args):
 
             for model in models:
                 if isinstance(model, type(accelerator.unwrap_model(unet))):
-                    unet_lora_layers_to_save = get_peft_model_state_dict(model)
+                    unet_lora_layers_to_save = unet_lora_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder))):
-                    text_encoder_lora_layers_to_save = get_peft_model_state_dict(model)
+                    text_encoder_lora_layers_to_save = text_encoder_lora_state_dict(model)
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
@@ -944,10 +1010,11 @@ def main(args):
         optimizer_class = torch.optim.AdamW
 
     # Optimizer creation
-    params_to_optimize = list(filter(lambda p: p.requires_grad, unet.parameters()))
-    if args.train_text_encoder:
-        params_to_optimize = params_to_optimize + list(filter(lambda p: p.requires_grad, text_encoder.parameters()))
-
+    params_to_optimize = (
+        itertools.chain(unet_lora_parameters, text_lora_parameters)
+        if args.train_text_encoder
+        else unet_lora_parameters
+    )
     optimizer = optimizer_class(
         params_to_optimize,
         lr=args.learning_rate,
@@ -1190,7 +1257,12 @@ def main(args):
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    accelerator.clip_grad_norm_(params_to_optimize, args.max_grad_norm)
+                    params_to_clip = (
+                        itertools.chain(unet_lora_parameters, text_lora_parameters)
+                        if args.train_text_encoder
+                        else unet_lora_parameters
+                    )
+                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()
@@ -1313,19 +1385,19 @@ def main(args):
     if accelerator.is_main_process:
         unet = accelerator.unwrap_model(unet)
         unet = unet.to(torch.float32)
+        unet_lora_layers = unet_lora_state_dict(unet)
 
-        unet_lora_state_dict = get_peft_model_state_dict(unet)
-
-        if args.train_text_encoder:
+        if text_encoder is not None and args.train_text_encoder:
             text_encoder = accelerator.unwrap_model(text_encoder)
-            text_encoder_state_dict = get_peft_model_state_dict(text_encoder)
+            text_encoder = text_encoder.to(torch.float32)
+            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder)
         else:
-            text_encoder_state_dict = None
+            text_encoder_lora_layers = None
 
         LoraLoaderMixin.save_lora_weights(
             save_directory=args.output_dir,
-            unet_lora_layers=unet_lora_state_dict,
-            text_encoder_lora_layers=text_encoder_state_dict,
+            unet_lora_layers=unet_lora_layers,
+            text_encoder_lora_layers=text_encoder_lora_layers,
         )
 
         # Final inference

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -34,8 +34,6 @@ from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration
 from huggingface_hub import create_repo, upload_folder
 from huggingface_hub.utils import insecure_hashlib
 from packaging import version
-from peft import LoraConfig
-from peft.utils import get_peft_model_state_dict
 from PIL import Image
 from PIL.ImageOps import exif_transpose
 from torch.utils.data import Dataset
@@ -52,8 +50,9 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.loaders import LoraLoaderMixin
+from diffusers.models.lora import LoRALinearLayer
 from diffusers.optimization import get_scheduler
-from diffusers.training_utils import compute_snr
+from diffusers.training_utils import compute_snr, unet_lora_state_dict
 from diffusers.utils import check_min_version, is_wandb_available
 from diffusers.utils.import_utils import is_xformers_available
 
@@ -1010,19 +1009,54 @@ def main(args):
             text_encoder_two.gradient_checkpointing_enable()
 
     # now we will add new LoRA weights to the attention layers
-    unet_lora_config = LoraConfig(
-        r=args.rank, init_lora_weights="gaussian", target_modules=["to_k", "to_q", "to_v", "to_out.0"]
-    )
-    unet.add_adapter(unet_lora_config)
+    # Set correct lora layers
+    unet_lora_parameters = []
+    for attn_processor_name, attn_processor in unet.attn_processors.items():
+        # Parse the attention module.
+        attn_module = unet
+        for n in attn_processor_name.split(".")[:-1]:
+            attn_module = getattr(attn_module, n)
+
+        # Set the `lora_layer` attribute of the attention-related matrices.
+        attn_module.to_q.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_k.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_v.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_out[0].set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_out[0].in_features,
+                out_features=attn_module.to_out[0].out_features,
+                rank=args.rank,
+            )
+        )
+
+        # Accumulate the LoRA params to optimize.
+        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
 
     # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
     # So, instead, we monkey-patch the forward calls of its attention-blocks.
     if args.train_text_encoder:
-        text_lora_config = LoraConfig(
-            r=args.rank, init_lora_weights="gaussian", target_modules=["q_proj", "k_proj", "v_proj", "out_proj"]
+        # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
+        text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder(
+            text_encoder_one, dtype=torch.float32, rank=args.rank
+        )
+        text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder(
+            text_encoder_two, dtype=torch.float32, rank=args.rank
         )
-        text_encoder_one.add_adapter(text_lora_config)
-        text_encoder_two.add_adapter(text_lora_config)
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
@@ -1035,11 +1069,11 @@ def main(args):
 
             for model in models:
                 if isinstance(model, type(accelerator.unwrap_model(unet))):
-                    unet_lora_layers_to_save = get_peft_model_state_dict(model)
+                    unet_lora_layers_to_save = unet_lora_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
-                    text_encoder_one_lora_layers_to_save = get_peft_model_state_dict(model)
+                    text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
-                    text_encoder_two_lora_layers_to_save = get_peft_model_state_dict(model)
+                    text_encoder_two_lora_layers_to_save = text_encoder_lora_state_dict(model)
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
@@ -1096,12 +1130,6 @@ def main(args):
             args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
         )
 
-    unet_lora_parameters = list(filter(lambda p: p.requires_grad, unet.parameters()))
-
-    if args.train_text_encoder:
-        text_lora_parameters_one = list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
-        text_lora_parameters_two = list(filter(lambda p: p.requires_grad, text_encoder_two.parameters()))
-
     # Optimization parameters
     unet_lora_parameters_with_lr = {"params": unet_lora_parameters, "lr": args.learning_rate}
     if args.train_text_encoder:
@@ -1166,10 +1194,26 @@ def main(args):
 
         optimizer_class = prodigyopt.Prodigy
 
+        if args.learning_rate <= 0.1:
+            logger.warn(
+                "Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
+            )
+        if args.train_text_encoder and args.text_encoder_lr:
+            logger.warn(
+                f"Learning rates were provided both for the unet and the text encoder- e.g. text_encoder_lr:"
+                f" {args.text_encoder_lr} and learning_rate: {args.learning_rate}. "
+                f"When using prodigy only learning_rate is used as the initial learning rate."
+            )
+            # changes the learning rate of text_encoder_parameters_one and text_encoder_parameters_two to be
+            # --learning_rate
+            params_to_optimize[1]["lr"] = args.learning_rate
+            params_to_optimize[2]["lr"] = args.learning_rate
+
         optimizer = optimizer_class(
             params_to_optimize,
             lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
+            beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,
             eps=args.adam_epsilon,
             decouple=args.prodigy_decouple,
@@ -1615,13 +1659,13 @@ def main(args):
     if accelerator.is_main_process:
         unet = accelerator.unwrap_model(unet)
         unet = unet.to(torch.float32)
-        unet_lora_layers = get_peft_model_state_dict(unet)
+        unet_lora_layers = unet_lora_state_dict(unet)
 
         if args.train_text_encoder:
             text_encoder_one = accelerator.unwrap_model(text_encoder_one)
-            text_encoder_lora_layers = get_peft_model_state_dict(text_encoder_one.to(torch.float32))
+            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder_one.to(torch.float32))
             text_encoder_two = accelerator.unwrap_model(text_encoder_two)
-            text_encoder_2_lora_layers = get_peft_model_state_dict(text_encoder_two.to(torch.float32))
+            text_encoder_2_lora_layers = text_encoder_lora_state_dict(text_encoder_two.to(torch.float32))
         else:
             text_encoder_lora_layers = None
             text_encoder_2_lora_layers = None

examples/research_projects/controlnet/train_controlnet_webdataset.py

@@ -420,7 +420,7 @@ def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
     text_encoder_config = PretrainedConfig.from_pretrained(
-        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision, use_auth_token=True
     )
     model_class = text_encoder_config.architectures[0]
 
@@ -975,7 +975,7 @@ def main(args):
         revision=args.revision,
     )
     unet = UNet2DConditionModel.from_pretrained(
-        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
+        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, use_auth_token=True
     )
 
     if args.controlnet_model_name_or_path:

examples/text_to_image/README.md

@@ -32,8 +32,6 @@ And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) e
 accelerate config
 ```
 
-Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
-
 ### Pokemon example
 
 You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. 

examples/text_to_image/README_sdxl.md

@@ -45,7 +45,6 @@ write_basic_config()
 ```
 
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
-Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
 
 ### Training
 

examples/text_to_image/requirements.txt

@@ -5,4 +5,3 @@ datasets
 ftfy
 tensorboard
 Jinja2
-peft==0.7.0

examples/text_to_image/requirements_sdxl.txt

@@ -5,4 +5,3 @@ ftfy
 tensorboard
 Jinja2
 datasets
-peft==0.7.0

examples/text_to_image/train_text_to_image_lora.py

@@ -34,14 +34,13 @@ from accelerate.utils import ProjectConfiguration, set_seed
 from datasets import load_dataset
 from huggingface_hub import create_repo, upload_folder
 from packaging import version
-from peft import LoraConfig
-from peft.utils import get_peft_model_state_dict
 from torchvision import transforms
 from tqdm.auto import tqdm
 from transformers import CLIPTextModel, CLIPTokenizer
 
 import diffusers
-from diffusers import AutoencoderKL, DDPMScheduler, DiffusionPipeline, StableDiffusionPipeline, UNet2DConditionModel
+from diffusers import AutoencoderKL, DDPMScheduler, DiffusionPipeline, UNet2DConditionModel
+from diffusers.models.lora import LoRALinearLayer
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_snr
 from diffusers.utils import check_min_version, is_wandb_available
@@ -480,20 +479,62 @@ def main():
     elif accelerator.mixed_precision == "bf16":
         weight_dtype = torch.bfloat16
 
-    # Freeze the unet parameters before adding adapters
-    for param in unet.parameters():
-        param.requires_grad_(False)
-
-    unet_lora_config = LoraConfig(
-        r=args.rank, init_lora_weights="gaussian", target_modules=["to_k", "to_q", "to_v", "to_out.0"]
-    )
-
     # Move unet, vae and text_encoder to device and cast to weight_dtype
     unet.to(accelerator.device, dtype=weight_dtype)
     vae.to(accelerator.device, dtype=weight_dtype)
     text_encoder.to(accelerator.device, dtype=weight_dtype)
 
-    unet.add_adapter(unet_lora_config)
+    # now we will add new LoRA weights to the attention layers
+    # It's important to realize here how many attention weights will be added and of which sizes
+    # The sizes of the attention layers consist only of two different variables:
+    # 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.
+    # 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.
+
+    # Let's first see how many attention processors we will have to set.
+    # For Stable Diffusion, it should be equal to:
+    # - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12
+    # - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2
+    # - up blocks (2x attention layers) * (3x transformer layers) * (3x down blocks) = 18
+    # => 32 layers
+
+    # Set correct lora layers
+    unet_lora_parameters = []
+    for attn_processor_name, attn_processor in unet.attn_processors.items():
+        # Parse the attention module.
+        attn_module = unet
+        for n in attn_processor_name.split(".")[:-1]:
+            attn_module = getattr(attn_module, n)
+
+        # Set the `lora_layer` attribute of the attention-related matrices.
+        attn_module.to_q.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_k.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
+            )
+        )
+
+        attn_module.to_v.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_out[0].set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_out[0].in_features,
+                out_features=attn_module.to_out[0].out_features,
+                rank=args.rank,
+            )
+        )
+
+        # Accumulate the LoRA params to optimize.
+        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
 
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
@@ -508,8 +549,6 @@ def main():
         else:
             raise ValueError("xformers is not available. Make sure it is installed correctly")
 
-    lora_layers = filter(lambda p: p.requires_grad, unet.parameters())
-
     # Enable TF32 for faster training on Ampere GPUs,
     # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
     if args.allow_tf32:
@@ -534,7 +573,7 @@ def main():
         optimizer_cls = torch.optim.AdamW
 
     optimizer = optimizer_cls(
-        lora_layers,
+        unet_lora_parameters,
         lr=args.learning_rate,
         betas=(args.adam_beta1, args.adam_beta2),
         weight_decay=args.adam_weight_decay,
@@ -661,8 +700,8 @@ def main():
     )
 
     # Prepare everything with our `accelerator`.
-    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        unet, optimizer, train_dataloader, lr_scheduler
+    unet_lora_parameters, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        unet_lora_parameters, optimizer, train_dataloader, lr_scheduler
     )
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
@@ -794,7 +833,7 @@ def main():
                 # Backpropagate
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    params_to_clip = lora_layers
+                    params_to_clip = unet_lora_parameters
                     accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
@@ -831,15 +870,6 @@ def main():
 
                         save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                         accelerator.save_state(save_path)
-
-                        unet_lora_state_dict = get_peft_model_state_dict(unet)
-
-                        StableDiffusionPipeline.save_lora_weights(
-                            save_directory=save_path,
-                            unet_lora_layers=unet_lora_state_dict,
-                            safe_serialization=True,
-                        )
-
                         logger.info(f"Saved state to {save_path}")
 
             logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
@@ -896,13 +926,7 @@ def main():
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         unet = unet.to(torch.float32)
-
-        unet_lora_state_dict = get_peft_model_state_dict(unet)
-        StableDiffusionPipeline.save_lora_weights(
-            save_directory=args.output_dir,
-            unet_lora_layers=unet_lora_state_dict,
-            safe_serialization=True,
-        )
+        unet.save_attn_procs(args.output_dir)
 
         if args.push_to_hub:
             save_model_card(

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -16,6 +16,7 @@
 """Fine-tuning script for Stable Diffusion XL for text2image with support for LoRA."""
 
 import argparse
+import itertools
 import logging
 import math
 import os
@@ -36,8 +37,6 @@ from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration
 from datasets import load_dataset
 from huggingface_hub import create_repo, upload_folder
 from packaging import version
-from peft import LoraConfig
-from peft.utils import get_peft_model_state_dict
 from torchvision import transforms
 from torchvision.transforms.functional import crop
 from tqdm.auto import tqdm
@@ -51,6 +50,7 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.loaders import LoraLoaderMixin
+from diffusers.models.lora import LoRALinearLayer
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_snr
 from diffusers.utils import check_min_version, is_wandb_available
@@ -658,20 +658,53 @@ def main(args):
 
     # now we will add new LoRA weights to the attention layers
     # Set correct lora layers
-    unet_lora_config = LoraConfig(
-        r=args.rank, init_lora_weights="gaussian", target_modules=["to_k", "to_q", "to_v", "to_out.0"]
-    )
+    unet_lora_parameters = []
+    for attn_processor_name, attn_processor in unet.attn_processors.items():
+        # Parse the attention module.
+        attn_module = unet
+        for n in attn_processor_name.split(".")[:-1]:
+            attn_module = getattr(attn_module, n)
 
-    unet.add_adapter(unet_lora_config)
+        # Set the `lora_layer` attribute of the attention-related matrices.
+        attn_module.to_q.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_k.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_v.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_out[0].set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_out[0].in_features,
+                out_features=attn_module.to_out[0].out_features,
+                rank=args.rank,
+            )
+        )
 
-    # The text encoder comes from 🤗 transformers, we will also attach adapters to it.
+        # Accumulate the LoRA params to optimize.
+        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
+
+    # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
+    # So, instead, we monkey-patch the forward calls of its attention-blocks.
     if args.train_text_encoder:
         # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
-        text_lora_config = LoraConfig(
-            r=args.rank, init_lora_weights="gaussian", target_modules=["q_proj", "k_proj", "v_proj", "out_proj"]
+        text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder(
+            text_encoder_one, dtype=torch.float32, rank=args.rank
+        )
+        text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder(
+            text_encoder_two, dtype=torch.float32, rank=args.rank
         )
-        text_encoder_one.add_adapter(text_lora_config)
-        text_encoder_two.add_adapter(text_lora_config)
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
@@ -684,11 +717,11 @@ def main(args):
 
             for model in models:
                 if isinstance(model, type(accelerator.unwrap_model(unet))):
-                    unet_lora_layers_to_save = get_peft_model_state_dict(model)
+                    unet_lora_layers_to_save = unet_attn_processors_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
-                    text_encoder_one_lora_layers_to_save = get_peft_model_state_dict(model)
+                    text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(model)
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
-                    text_encoder_two_lora_layers_to_save = get_peft_model_state_dict(model)
+                    text_encoder_two_lora_layers_to_save = text_encoder_lora_state_dict(model)
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
@@ -759,13 +792,11 @@ def main(args):
         optimizer_class = torch.optim.AdamW
 
     # Optimizer creation
-    params_to_optimize = list(filter(lambda p: p.requires_grad, unet.parameters()))
-    if args.train_text_encoder:
-        params_to_optimize = (
-            params_to_optimize
-            + list(filter(lambda p: p.requires_grad, text_encoder_one.parameters()))
-            + list(filter(lambda p: p.requires_grad, text_encoder_two.parameters()))
-        )
+    params_to_optimize = (
+        itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
+        if args.train_text_encoder
+        else unet_lora_parameters
+    )
     optimizer = optimizer_class(
         params_to_optimize,
         lr=args.learning_rate,
@@ -1097,7 +1128,12 @@ def main(args):
                 # Backpropagate
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    accelerator.clip_grad_norm_(params_to_optimize, args.max_grad_norm)
+                    params_to_clip = (
+                        itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
+                        if args.train_text_encoder
+                        else unet_lora_parameters
+                    )
+                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()
@@ -1193,21 +1229,20 @@ def main(args):
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         unet = accelerator.unwrap_model(unet)
-        unet_lora_state_dict = get_peft_model_state_dict(unet)
+        unet_lora_layers = unet_attn_processors_state_dict(unet)
 
         if args.train_text_encoder:
             text_encoder_one = accelerator.unwrap_model(text_encoder_one)
+            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder_one)
             text_encoder_two = accelerator.unwrap_model(text_encoder_two)
-
-            text_encoder_lora_layers = get_peft_model_state_dict(text_encoder_one)
-            text_encoder_2_lora_layers = get_peft_model_state_dict(text_encoder_two)
+            text_encoder_2_lora_layers = text_encoder_lora_state_dict(text_encoder_two)
         else:
             text_encoder_lora_layers = None
             text_encoder_2_lora_layers = None
 
         StableDiffusionXLPipeline.save_lora_weights(
             save_directory=args.output_dir,
-            unet_lora_layers=unet_lora_state_dict,
+            unet_lora_layers=unet_lora_layers,
             text_encoder_lora_layers=text_encoder_lora_layers,
             text_encoder_2_lora_layers=text_encoder_2_lora_layers,
         )

src/diffusers/__init__.py

@@ -80,7 +80,6 @@ else:
             "AutoencoderTiny",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
-            "ControlNetXSModel",
             "Kandinsky3UNet",
             "ModelMixin",
             "MotionAdapter",
@@ -251,7 +250,6 @@ else:
             "StableDiffusionControlNetImg2ImgPipeline",
             "StableDiffusionControlNetInpaintPipeline",
             "StableDiffusionControlNetPipeline",
-            "StableDiffusionControlNetXSPipeline",
             "StableDiffusionDepth2ImgPipeline",
             "StableDiffusionDiffEditPipeline",
             "StableDiffusionGLIGENPipeline",
@@ -275,7 +273,6 @@ else:
             "StableDiffusionXLControlNetImg2ImgPipeline",
             "StableDiffusionXLControlNetInpaintPipeline",
             "StableDiffusionXLControlNetPipeline",
-            "StableDiffusionXLControlNetXSPipeline",
             "StableDiffusionXLImg2ImgPipeline",
             "StableDiffusionXLInpaintPipeline",
             "StableDiffusionXLInstructPix2PixPipeline",
@@ -457,7 +454,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderTiny,
             ConsistencyDecoderVAE,
             ControlNetModel,
-            ControlNetXSModel,
             Kandinsky3UNet,
             ModelMixin,
             MotionAdapter,
@@ -607,7 +603,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionControlNetImg2ImgPipeline,
             StableDiffusionControlNetInpaintPipeline,
             StableDiffusionControlNetPipeline,
-            StableDiffusionControlNetXSPipeline,
             StableDiffusionDepth2ImgPipeline,
             StableDiffusionDiffEditPipeline,
             StableDiffusionGLIGENPipeline,
@@ -631,7 +626,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionXLControlNetImg2ImgPipeline,
             StableDiffusionXLControlNetInpaintPipeline,
             StableDiffusionXLControlNetPipeline,
-            StableDiffusionXLControlNetXSPipeline,
             StableDiffusionXLImg2ImgPipeline,
             StableDiffusionXLInpaintPipeline,
             StableDiffusionXLInstructPix2PixPipeline,

src/diffusers/commands/fp16_safetensors.py

@@ -19,7 +19,6 @@ Usage example:
 
 import glob
 import json
-import warnings
 from argparse import ArgumentParser, Namespace
 from importlib import import_module
 
@@ -33,12 +32,12 @@ from . import BaseDiffusersCLICommand
 
 
 def conversion_command_factory(args: Namespace):
-    if args.use_auth_token:
-        warnings.warn(
-            "The `--use_auth_token` flag is deprecated and will be removed in a future version. Authentication is now"
-            " handled automatically if user is logged in."
-        )
-    return FP16SafetensorsCommand(args.ckpt_id, args.fp16, args.use_safetensors)
+    return FP16SafetensorsCommand(
+        args.ckpt_id,
+        args.fp16,
+        args.use_safetensors,
+        args.use_auth_token,
+    )
 
 
 class FP16SafetensorsCommand(BaseDiffusersCLICommand):
@@ -63,7 +62,7 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
         )
         conversion_parser.set_defaults(func=conversion_command_factory)
 
-    def __init__(self, ckpt_id: str, fp16: bool, use_safetensors: bool):
+    def __init__(self, ckpt_id: str, fp16: bool, use_safetensors: bool, use_auth_token: bool):
         self.logger = logging.get_logger("diffusers-cli/fp16_safetensors")
         self.ckpt_id = ckpt_id
         self.local_ckpt_dir = f"/tmp/{ckpt_id}"
@@ -76,6 +75,8 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
                 "When `use_safetensors` and `fp16` both are False, then this command is of no use."
             )
 
+        self.use_auth_token = use_auth_token
+
     def run(self):
         if version.parse(huggingface_hub.__version__) < version.parse("0.9.0"):
             raise ImportError(
@@ -86,7 +87,7 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
             from huggingface_hub import create_commit
             from huggingface_hub._commit_api import CommitOperationAdd
 
-        model_index = hf_hub_download(repo_id=self.ckpt_id, filename="model_index.json")
+        model_index = hf_hub_download(repo_id=self.ckpt_id, filename="model_index.json", token=self.use_auth_token)
         with open(model_index, "r") as f:
             pipeline_class_name = json.load(f)["_class_name"]
         pipeline_class = getattr(import_module("diffusers"), pipeline_class_name)
@@ -95,7 +96,7 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
         # Load the appropriate pipeline. We could have use `DiffusionPipeline`
         # here, but just to avoid any rough edge cases.
         pipeline = pipeline_class.from_pretrained(
-            self.ckpt_id, torch_dtype=torch.float16 if self.fp16 else torch.float32
+            self.ckpt_id, torch_dtype=torch.float16 if self.fp16 else torch.float32, use_auth_token=self.use_auth_token
         )
         pipeline.save_pretrained(
             self.local_ckpt_dir,

src/diffusers/configuration_utils.py

@@ -27,16 +27,12 @@ from typing import Any, Dict, Tuple, Union
 
 import numpy as np
 from huggingface_hub import create_repo, hf_hub_download
-from huggingface_hub.utils import (
-    EntryNotFoundError,
-    RepositoryNotFoundError,
-    RevisionNotFoundError,
-    validate_hf_hub_args,
-)
+from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
 from requests import HTTPError
 
 from . import __version__
 from .utils import (
+    DIFFUSERS_CACHE,
     HUGGINGFACE_CO_RESOLVE_ENDPOINT,
     DummyObject,
     deprecate,
@@ -279,7 +275,6 @@ class ConfigMixin:
         return cls.load_config(*args, **kwargs)
 
     @classmethod
-    @validate_hf_hub_args
     def load_config(
         cls,
         pretrained_model_name_or_path: Union[str, os.PathLike],
@@ -316,7 +311,7 @@ class ConfigMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -334,11 +329,11 @@ class ConfigMixin:
                 A dictionary of all the parameters stored in a JSON configuration file.
 
         """
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
         _ = kwargs.pop("mirror", None)
@@ -381,7 +376,7 @@ class ConfigMixin:
                     proxies=proxies,
                     resume_download=resume_download,
                     local_files_only=local_files_only,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     user_agent=user_agent,
                     subfolder=subfolder,
                     revision=revision,
@@ -390,7 +385,8 @@ class ConfigMixin:
                 raise EnvironmentError(
                     f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier"
                     " listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a"
-                    " token having permission to this repo with `token` or log in with `huggingface-cli login`."
+                    " token having permission to this repo with `use_auth_token` or log in with `huggingface-cli"
+                    " login`."
                 )
             except RevisionNotFoundError:
                 raise EnvironmentError(

src/diffusers/loaders/ip_adapter.py

@@ -15,10 +15,11 @@ import os
 from typing import Dict, Union
 
 import torch
-from huggingface_hub.utils import validate_hf_hub_args
 from safetensors import safe_open
 
 from ..utils import (
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
     _get_model_file,
     is_transformers_available,
     logging,
@@ -42,7 +43,6 @@ logger = logging.get_logger(__name__)
 class IPAdapterMixin:
     """Mixin for handling IP Adapters."""
 
-    @validate_hf_hub_args
     def load_ip_adapter(
         self,
         pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
@@ -77,7 +77,7 @@ class IPAdapterMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -88,12 +88,12 @@ class IPAdapterMixin:
         """
 
         # Load the main state dict first.
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
 
         user_agent = {
@@ -110,7 +110,7 @@ class IPAdapterMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 subfolder=subfolder,
                 user_agent=user_agent,

src/diffusers/loaders/lora.py

@@ -18,13 +18,14 @@ from typing import Callable, Dict, List, Optional, Union
 import safetensors
 import torch
 from huggingface_hub import model_info
-from huggingface_hub.utils import validate_hf_hub_args
 from packaging import version
 from torch import nn
 
 from .. import __version__
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import (
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
     USE_PEFT_BACKEND,
     _get_model_file,
     convert_state_dict_to_diffusers,
@@ -131,7 +132,6 @@ class LoraLoaderMixin:
         )
 
     @classmethod
-    @validate_hf_hub_args
     def lora_state_dict(
         cls,
         pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
@@ -174,7 +174,7 @@ class LoraLoaderMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -195,12 +195,12 @@ class LoraLoaderMixin:
         """
         # Load the main state dict first which has the LoRA layers for either of
         # UNet and text encoder or both.
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         subfolder = kwargs.pop("subfolder", None)
         weight_name = kwargs.pop("weight_name", None)
@@ -239,7 +239,7 @@ class LoraLoaderMixin:
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        token=token,
+                        use_auth_token=use_auth_token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -265,7 +265,7 @@ class LoraLoaderMixin:
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,

src/diffusers/loaders/single_file.py

@@ -18,9 +18,10 @@ from pathlib import Path
 import requests
 import torch
 from huggingface_hub import hf_hub_download
-from huggingface_hub.utils import validate_hf_hub_args
 
 from ..utils import (
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
     deprecate,
     is_accelerate_available,
     is_omegaconf_available,
@@ -51,7 +52,6 @@ class FromSingleFileMixin:
         return cls.from_single_file(*args, **kwargs)
 
     @classmethod
-    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`DiffusionPipeline`] from pretrained pipeline weights saved in the `.ckpt` or `.safetensors`
@@ -81,7 +81,7 @@ class FromSingleFileMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -154,12 +154,12 @@ class FromSingleFileMixin:
 
         original_config_file = kwargs.pop("original_config_file", None)
         config_files = kwargs.pop("config_files", None)
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         extract_ema = kwargs.pop("extract_ema", False)
         image_size = kwargs.pop("image_size", None)
@@ -253,7 +253,7 @@ class FromSingleFileMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 force_download=force_download,
             )
@@ -282,7 +282,7 @@ class FromSingleFileMixin:
         )
 
         if torch_dtype is not None:
-            pipe.to(dtype=torch_dtype)
+            pipe.to(torch_dtype=torch_dtype)
 
         return pipe
 
@@ -293,7 +293,6 @@ class FromOriginalVAEMixin:
     """
 
     @classmethod
-    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`AutoencoderKL`] from pretrained ControlNet weights saved in the original `.ckpt` or
@@ -323,7 +322,7 @@ class FromOriginalVAEMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to True, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -380,12 +379,12 @@ class FromOriginalVAEMixin:
         )
 
         config_file = kwargs.pop("config_file", None)
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         image_size = kwargs.pop("image_size", None)
         scaling_factor = kwargs.pop("scaling_factor", None)
@@ -426,7 +425,7 @@ class FromOriginalVAEMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 force_download=force_download,
             )
@@ -491,7 +490,6 @@ class FromOriginalControlnetMixin:
     """
 
     @classmethod
-    @validate_hf_hub_args
     def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
         r"""
         Instantiate a [`ControlNetModel`] from pretrained ControlNet weights saved in the original `.ckpt` or
@@ -521,7 +519,7 @@ class FromOriginalControlnetMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to True, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -557,12 +555,12 @@ class FromOriginalControlnetMixin:
         from ..pipelines.stable_diffusion.convert_from_ckpt import download_controlnet_from_original_ckpt
 
         config_file = kwargs.pop("config_file", None)
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         num_in_channels = kwargs.pop("num_in_channels", None)
         use_linear_projection = kwargs.pop("use_linear_projection", None)
         revision = kwargs.pop("revision", None)
@@ -605,7 +603,7 @@ class FromOriginalControlnetMixin:
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 force_download=force_download,
             )

src/diffusers/loaders/textual_inversion.py

@@ -15,10 +15,16 @@ from typing import Dict, List, Optional, Union
 
 import safetensors
 import torch
-from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
-from ..utils import _get_model_file, is_accelerate_available, is_transformers_available, logging
+from ..utils import (
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
+    _get_model_file,
+    is_accelerate_available,
+    is_transformers_available,
+    logging,
+)
 
 
 if is_transformers_available():
@@ -33,14 +39,13 @@ TEXT_INVERSION_NAME = "learned_embeds.bin"
 TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
 
 
-@validate_hf_hub_args
 def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
-    cache_dir = kwargs.pop("cache_dir", None)
+    cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
     force_download = kwargs.pop("force_download", False)
     resume_download = kwargs.pop("resume_download", False)
     proxies = kwargs.pop("proxies", None)
-    local_files_only = kwargs.pop("local_files_only", None)
-    token = kwargs.pop("token", None)
+    local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+    use_auth_token = kwargs.pop("use_auth_token", None)
     revision = kwargs.pop("revision", None)
     subfolder = kwargs.pop("subfolder", None)
     weight_name = kwargs.pop("weight_name", None)
@@ -74,7 +79,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        token=token,
+                        use_auth_token=use_auth_token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -95,7 +100,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,
@@ -262,7 +267,6 @@ class TextualInversionLoaderMixin:
 
         return all_tokens, all_embeddings
 
-    @validate_hf_hub_args
     def load_textual_inversion(
         self,
         pretrained_model_name_or_path: Union[str, List[str], Dict[str, torch.Tensor], List[Dict[str, torch.Tensor]]],
@@ -316,7 +320,7 @@ class TextualInversionLoaderMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):

src/diffusers/loaders/unet.py

@@ -19,12 +19,13 @@ from typing import Callable, Dict, List, Optional, Union
 import safetensors
 import torch
 import torch.nn.functional as F
-from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
-from ..models.embeddings import ImageProjection, MLPProjection, Resampler
+from ..models.embeddings import ImageProjection, Resampler
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import (
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
     USE_PEFT_BACKEND,
     _get_model_file,
     delete_adapter_layers,
@@ -61,7 +62,6 @@ class UNet2DConditionLoadersMixin:
     text_encoder_name = TEXT_ENCODER_NAME
     unet_name = UNET_NAME
 
-    @validate_hf_hub_args
     def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
         r"""
         Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
@@ -95,7 +95,7 @@ class UNet2DConditionLoadersMixin:
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
@@ -130,12 +130,12 @@ class UNet2DConditionLoadersMixin:
         from ..models.attention_processor import CustomDiffusionAttnProcessor
         from ..models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
 
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         subfolder = kwargs.pop("subfolder", None)
         weight_name = kwargs.pop("weight_name", None)
@@ -184,7 +184,7 @@ class UNet2DConditionLoadersMixin:
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        token=token,
+                        use_auth_token=use_auth_token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -204,7 +204,7 @@ class UNet2DConditionLoadersMixin:
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,
@@ -675,9 +675,6 @@ class UNet2DConditionLoadersMixin:
         if "proj.weight" in state_dict["image_proj"]:
             # IP-Adapter
             num_image_text_embeds = 4
-        elif "proj.3.weight" in state_dict["image_proj"]:
-            # IP-Adapter Full Face
-            num_image_text_embeds = 257  # 256 CLIP tokens + 1 CLS token
         else:
             # IP-Adapter Plus
             num_image_text_embeds = state_dict["image_proj"]["latents"].shape[1]
@@ -747,32 +744,8 @@ class UNet2DConditionLoadersMixin:
                     "norm.bias": state_dict["image_proj"]["norm.bias"],
                 }
             )
+
             image_projection.load_state_dict(image_proj_state_dict)
-            del image_proj_state_dict
-
-        elif "proj.3.weight" in state_dict["image_proj"]:
-            clip_embeddings_dim = state_dict["image_proj"]["proj.0.weight"].shape[0]
-            cross_attention_dim = state_dict["image_proj"]["proj.3.weight"].shape[0]
-
-            image_projection = MLPProjection(
-                cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim
-            )
-            image_projection.to(dtype=self.dtype, device=self.device)
-
-            # load image projection layer weights
-            image_proj_state_dict = {}
-            image_proj_state_dict.update(
-                {
-                    "ff.net.0.proj.weight": state_dict["image_proj"]["proj.0.weight"],
-                    "ff.net.0.proj.bias": state_dict["image_proj"]["proj.0.bias"],
-                    "ff.net.2.weight": state_dict["image_proj"]["proj.2.weight"],
-                    "ff.net.2.bias": state_dict["image_proj"]["proj.2.bias"],
-                    "norm.weight": state_dict["image_proj"]["proj.3.weight"],
-                    "norm.bias": state_dict["image_proj"]["proj.3.bias"],
-                }
-            )
-            image_projection.load_state_dict(image_proj_state_dict)
-            del image_proj_state_dict
 
         else:
             # IP-Adapter Plus

src/diffusers/models/__init__.py

@@ -32,10 +32,9 @@ if is_torch_available():
     _import_structure["autoencoder_tiny"] = ["AutoencoderTiny"]
     _import_structure["consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["controlnet"] = ["ControlNetModel"]
-    _import_structure["controlnetxs"] = ["ControlNetXSModel"]
     _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
-    _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
+    _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["prior_transformer"] = ["PriorTransformer"]
     _import_structure["t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformer_2d"] = ["Transformer2DModel"]
@@ -64,7 +63,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .autoencoder_tiny import AutoencoderTiny
         from .consistency_decoder_vae import ConsistencyDecoderVAE
         from .controlnet import ControlNetModel
-        from .controlnetxs import ControlNetXSModel
         from .dual_transformer_2d import DualTransformer2DModel
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin

src/diffusers/models/attention_processor.py

@@ -113,14 +113,12 @@ class Attention(nn.Module):
     ):
         super().__init__()
         self.inner_dim = out_dim if out_dim is not None else dim_head * heads
-        self.query_dim = query_dim
         self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
         self.upcast_attention = upcast_attention
         self.upcast_softmax = upcast_softmax
         self.rescale_output_factor = rescale_output_factor
         self.residual_connection = residual_connection
         self.dropout = dropout
-        self.fused_projections = False
         self.out_dim = out_dim if out_dim is not None else query_dim
 
         # we make use of this private variable to know whether this class is loaded
@@ -182,7 +180,6 @@ class Attention(nn.Module):
         else:
             linear_cls = LoRACompatibleLinear
 
-        self.linear_cls = linear_cls
         self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias)
 
         if not self.only_cross_attention:
@@ -695,32 +692,6 @@ class Attention(nn.Module):
 
         return encoder_hidden_states
 
-    @torch.no_grad()
-    def fuse_projections(self, fuse=True):
-        is_cross_attention = self.cross_attention_dim != self.query_dim
-        device = self.to_q.weight.data.device
-        dtype = self.to_q.weight.data.dtype
-
-        if not is_cross_attention:
-            # fetch weight matrices.
-            concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
-            in_features = concatenated_weights.shape[1]
-            out_features = concatenated_weights.shape[0]
-
-            # create a new single projection layer and copy over the weights.
-            self.to_qkv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype)
-            self.to_qkv.weight.copy_(concatenated_weights)
-
-        else:
-            concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
-            in_features = concatenated_weights.shape[1]
-            out_features = concatenated_weights.shape[0]
-
-            self.to_kv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype)
-            self.to_kv.weight.copy_(concatenated_weights)
-
-        self.fused_projections = fuse
-
 
 class AttnProcessor:
     r"""
@@ -1213,6 +1184,9 @@ class AttnProcessor2_0:
         scale: float = 1.0,
     ) -> torch.FloatTensor:
         residual = hidden_states
+
+        args = () if USE_PEFT_BACKEND else (scale,)
+
         if attn.spatial_norm is not None:
             hidden_states = attn.spatial_norm(hidden_states, temb)
 
@@ -1279,103 +1253,6 @@ class AttnProcessor2_0:
         return hidden_states
 
 
-class FusedAttnProcessor2_0:
-    r"""
-    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
-    It uses fused projection layers. For self-attention modules, all projection matrices (i.e., query,
-    key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
-
-    <Tip warning={true}>
-
-    This API is currently 🧪 experimental in nature and can change in future.
-
-    </Tip>
-    """
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError(
-                "FusedAttnProcessor2_0 requires at least PyTorch 2.0, to use it. Please upgrade PyTorch to > 2.0."
-            )
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.FloatTensor,
-        encoder_hidden_states: Optional[torch.FloatTensor] = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        temb: Optional[torch.FloatTensor] = None,
-        scale: float = 1.0,
-    ) -> torch.FloatTensor:
-        residual = hidden_states
-        if attn.spatial_norm is not None:
-            hidden_states = attn.spatial_norm(hidden_states, temb)
-
-        input_ndim = hidden_states.ndim
-
-        if input_ndim == 4:
-            batch_size, channel, height, width = hidden_states.shape
-            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
-
-        batch_size, sequence_length, _ = (
-            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-        )
-
-        if attention_mask is not None:
-            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
-            # scaled_dot_product_attention expects attention_mask shape to be
-            # (batch, heads, source_length, target_length)
-            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
-
-        if attn.group_norm is not None:
-            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
-
-        args = () if USE_PEFT_BACKEND else (scale,)
-        if encoder_hidden_states is None:
-            qkv = attn.to_qkv(hidden_states, *args)
-            split_size = qkv.shape[-1] // 3
-            query, key, value = torch.split(qkv, split_size, dim=-1)
-        else:
-            if attn.norm_cross:
-                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
-            query = attn.to_q(hidden_states, *args)
-
-            kv = attn.to_kv(encoder_hidden_states, *args)
-            split_size = kv.shape[-1] // 2
-            key, value = torch.split(kv, split_size, dim=-1)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        # the output of sdp = (batch, num_heads, seq_len, head_dim)
-        # TODO: add support for attn.scale when we move to Torch 2.1
-        hidden_states = F.scaled_dot_product_attention(
-            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
-        )
-
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-        hidden_states = hidden_states.to(query.dtype)
-
-        # linear proj
-        hidden_states = attn.to_out[0](hidden_states, *args)
-        # dropout
-        hidden_states = attn.to_out[1](hidden_states)
-
-        if input_ndim == 4:
-            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
-
-        if attn.residual_connection:
-            hidden_states = hidden_states + residual
-
-        hidden_states = hidden_states / attn.rescale_output_factor
-
-        return hidden_states
-
-
 class CustomDiffusionXFormersAttnProcessor(nn.Module):
     r"""
     Processor for implementing memory efficient attention using xFormers for the Custom Diffusion method.
@@ -2374,7 +2251,6 @@ CROSS_ATTENTION_PROCESSORS = (
 AttentionProcessor = Union[
     AttnProcessor,
     AttnProcessor2_0,
-    FusedAttnProcessor2_0,
     XFormersAttnProcessor,
     SlicedAttnProcessor,
     AttnAddedKVProcessor,

src/diffusers/models/autoencoder_kl.py

@@ -22,7 +22,6 @@ from ..utils.accelerate_utils import apply_forward_hook
 from .attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
-    Attention,
     AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
@@ -449,41 +448,3 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
             return (dec,)
 
         return DecoderOutput(sample=dec)
-
-    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
-    def fuse_qkv_projections(self):
-        """
-        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
-        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-        """
-        self.original_attn_processors = None
-
-        for _, attn_processor in self.attn_processors.items():
-            if "Added" in str(attn_processor.__class__.__name__):
-                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
-
-        self.original_attn_processors = self.attn_processors
-
-        for module in self.modules():
-            if isinstance(module, Attention):
-                module.fuse_projections(fuse=True)
-
-    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
-    def unfuse_qkv_projections(self):
-        """Disables the fused QKV projection if enabled.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        """
-        if self.original_attn_processors is not None:
-            self.set_attn_processor(self.original_attn_processors)

src/diffusers/models/controlnetxs.py

@@ -1,977 +0,0 @@
-# Copyright 2023 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import math
-from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Tuple, Union
-
-import torch
-import torch.utils.checkpoint
-from torch import nn
-from torch.nn import functional as F
-from torch.nn.modules.normalization import GroupNorm
-
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import BaseOutput, logging
-from .attention_processor import (
-    AttentionProcessor,
-)
-from .autoencoder_kl import AutoencoderKL
-from .lora import LoRACompatibleConv
-from .modeling_utils import ModelMixin
-from .unet_2d_blocks import (
-    CrossAttnDownBlock2D,
-    CrossAttnUpBlock2D,
-    DownBlock2D,
-    Downsample2D,
-    ResnetBlock2D,
-    Transformer2DModel,
-    UpBlock2D,
-    Upsample2D,
-)
-from .unet_2d_condition import UNet2DConditionModel
-
-
-logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
-
-
-@dataclass
-class ControlNetXSOutput(BaseOutput):
-    """
-    The output of [`ControlNetXSModel`].
-
-    Args:
-        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
-            The output of the `ControlNetXSModel`. Unlike `ControlNetOutput` this is NOT to be added to the base model
-            output, but is already the final output.
-    """
-
-    sample: torch.FloatTensor = None
-
-
-# copied from diffusers.models.controlnet.ControlNetConditioningEmbedding
-class ControlNetConditioningEmbedding(nn.Module):
-    """
-    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
-    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
-    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
-    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
-    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
-    model) to encode image-space conditions ... into feature maps ..."
-    """
-
-    def __init__(
-        self,
-        conditioning_embedding_channels: int,
-        conditioning_channels: int = 3,
-        block_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
-    ):
-        super().__init__()
-
-        self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
-
-        self.blocks = nn.ModuleList([])
-
-        for i in range(len(block_out_channels) - 1):
-            channel_in = block_out_channels[i]
-            channel_out = block_out_channels[i + 1]
-            self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
-            self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
-
-        self.conv_out = zero_module(
-            nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
-        )
-
-    def forward(self, conditioning):
-        embedding = self.conv_in(conditioning)
-        embedding = F.silu(embedding)
-
-        for block in self.blocks:
-            embedding = block(embedding)
-            embedding = F.silu(embedding)
-
-        embedding = self.conv_out(embedding)
-
-        return embedding
-
-
-class ControlNetXSModel(ModelMixin, ConfigMixin):
-    r"""
-    A ControlNet-XS model
-
-    This model inherits from [`ModelMixin`] and [`ConfigMixin`]. Check the superclass documentation for it's generic
-    methods implemented for all models (such as downloading or saving).
-
-    Most of parameters for this model are passed into the [`UNet2DConditionModel`] it creates. Check the documentation
-    of [`UNet2DConditionModel`] for them.
-
-    Parameters:
-        conditioning_channels (`int`, defaults to 3):
-            Number of channels of conditioning input (e.g. an image)
-        controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
-            The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
-        conditioning_embedding_out_channels (`tuple[int]`, defaults to `(16, 32, 96, 256)`):
-            The tuple of output channel for each block in the `controlnet_cond_embedding` layer.
-        time_embedding_input_dim (`int`, defaults to 320):
-            Dimension of input into time embedding. Needs to be same as in the base model.
-        time_embedding_dim (`int`, defaults to 1280):
-            Dimension of output from time embedding. Needs to be same as in the base model.
-        learn_embedding (`bool`, defaults to `False`):
-            Whether to use time embedding of the control model. If yes, the time embedding is a linear interpolation of
-            the time embeddings of the control and base model with interpolation parameter `time_embedding_mix**3`.
-        time_embedding_mix (`float`, defaults to 1.0):
-            Linear interpolation parameter used if `learn_embedding` is `True`. A value of 1.0 means only the
-            control model's time embedding will be used. A value of 0.0 means only the base model's time embedding will be used.
-        base_model_channel_sizes (`Dict[str, List[Tuple[int]]]`):
-            Channel sizes of each subblock of base model. Use `gather_subblock_sizes` on your base model to compute it.
-    """
-
-    @classmethod
-    def init_original(cls, base_model: UNet2DConditionModel, is_sdxl=True):
-        """
-        Create a ControlNetXS model with the same parameters as in the original paper (https://github.com/vislearn/ControlNet-XS).
-
-        Parameters:
-            base_model (`UNet2DConditionModel`):
-                Base UNet model. Needs to be either StableDiffusion or StableDiffusion-XL.
-            is_sdxl (`bool`, defaults to `True`):
-                Whether passed `base_model` is a StableDiffusion-XL model.
-        """
-
-        def get_dim_attn_heads(base_model: UNet2DConditionModel, size_ratio: float, num_attn_heads: int):
-            """
-            Currently, diffusers can only set the dimension of attention heads (see https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 for why).
-            The original ControlNet-XS model, however, define the number of attention heads.
-            That's why compute the dimensions needed to get the correct number of attention heads.
-            """
-            block_out_channels = [int(size_ratio * c) for c in base_model.config.block_out_channels]
-            dim_attn_heads = [math.ceil(c / num_attn_heads) for c in block_out_channels]
-            return dim_attn_heads
-
-        if is_sdxl:
-            return ControlNetXSModel.from_unet(
-                base_model,
-                time_embedding_mix=0.95,
-                learn_embedding=True,
-                size_ratio=0.1,
-                conditioning_embedding_out_channels=(16, 32, 96, 256),
-                num_attention_heads=get_dim_attn_heads(base_model, 0.1, 64),
-            )
-        else:
-            return ControlNetXSModel.from_unet(
-                base_model,
-                time_embedding_mix=1.0,
-                learn_embedding=True,
-                size_ratio=0.0125,
-                conditioning_embedding_out_channels=(16, 32, 96, 256),
-                num_attention_heads=get_dim_attn_heads(base_model, 0.0125, 8),
-            )
-
-    @classmethod
-    def _gather_subblock_sizes(cls, unet: UNet2DConditionModel, base_or_control: str):
-        """To create correctly sized connections between base and control model, we need to know
-        the input and output channels of each subblock.
-
-        Parameters:
-            unet (`UNet2DConditionModel`):
-                Unet of which the subblock channels sizes are to be gathered.
-            base_or_control (`str`):
-                Needs to be either "base" or "control". If "base", decoder is also considered.
-        """
-        if base_or_control not in ["base", "control"]:
-            raise ValueError("`base_or_control` needs to be either `base` or `control`")
-
-        channel_sizes = {"down": [], "mid": [], "up": []}
-
-        # input convolution
-        channel_sizes["down"].append((unet.conv_in.in_channels, unet.conv_in.out_channels))
-
-        # encoder blocks
-        for module in unet.down_blocks:
-            if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
-                for r in module.resnets:
-                    channel_sizes["down"].append((r.in_channels, r.out_channels))
-                if module.downsamplers:
-                    channel_sizes["down"].append(
-                        (module.downsamplers[0].channels, module.downsamplers[0].out_channels)
-                    )
-            else:
-                raise ValueError(f"Encountered unknown module of type {type(module)} while creating ControlNet-XS.")
-
-        # middle block
-        channel_sizes["mid"].append((unet.mid_block.resnets[0].in_channels, unet.mid_block.resnets[0].out_channels))
-
-        # decoder blocks
-        if base_or_control == "base":
-            for module in unet.up_blocks:
-                if isinstance(module, (CrossAttnUpBlock2D, UpBlock2D)):
-                    for r in module.resnets:
-                        channel_sizes["up"].append((r.in_channels, r.out_channels))
-                else:
-                    raise ValueError(
-                        f"Encountered unknown module of type {type(module)} while creating ControlNet-XS."
-                    )
-
-        return channel_sizes
-
-    @register_to_config
-    def __init__(
-        self,
-        conditioning_channels: int = 3,
-        conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
-        controlnet_conditioning_channel_order: str = "rgb",
-        time_embedding_input_dim: int = 320,
-        time_embedding_dim: int = 1280,
-        time_embedding_mix: float = 1.0,
-        learn_embedding: bool = False,
-        base_model_channel_sizes: Dict[str, List[Tuple[int]]] = {
-            "down": [
-                (4, 320),
-                (320, 320),
-                (320, 320),
-                (320, 320),
-                (320, 640),
-                (640, 640),
-                (640, 640),
-                (640, 1280),
-                (1280, 1280),
-            ],
-            "mid": [(1280, 1280)],
-            "up": [
-                (2560, 1280),
-                (2560, 1280),
-                (1920, 1280),
-                (1920, 640),
-                (1280, 640),
-                (960, 640),
-                (960, 320),
-                (640, 320),
-                (640, 320),
-            ],
-        },
-        sample_size: Optional[int] = None,
-        down_block_types: Tuple[str] = (
-            "CrossAttnDownBlock2D",
-            "CrossAttnDownBlock2D",
-            "CrossAttnDownBlock2D",
-            "DownBlock2D",
-        ),
-        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
-        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
-        norm_num_groups: Optional[int] = 32,
-        cross_attention_dim: Union[int, Tuple[int]] = 1280,
-        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
-        num_attention_heads: Optional[Union[int, Tuple[int]]] = 8,
-        upcast_attention: bool = False,
-    ):
-        super().__init__()
-
-        # 1 - Create control unet
-        self.control_model = UNet2DConditionModel(
-            sample_size=sample_size,
-            down_block_types=down_block_types,
-            up_block_types=up_block_types,
-            block_out_channels=block_out_channels,
-            norm_num_groups=norm_num_groups,
-            cross_attention_dim=cross_attention_dim,
-            transformer_layers_per_block=transformer_layers_per_block,
-            attention_head_dim=num_attention_heads,
-            use_linear_projection=True,
-            upcast_attention=upcast_attention,
-            time_embedding_dim=time_embedding_dim,
-        )
-
-        # 2 - Do model surgery on control model
-        # 2.1 - Allow to use the same time information as the base model
-        adjust_time_dims(self.control_model, time_embedding_input_dim, time_embedding_dim)
-
-        # 2.2 - Allow for information infusion from base model
-
-        # We concat the output of each base encoder subblocks to the input of the next control encoder subblock
-        # (We ignore the 1st element, as it represents the `conv_in`.)
-        extra_input_channels = [input_channels for input_channels, _ in base_model_channel_sizes["down"][1:]]
-        it_extra_input_channels = iter(extra_input_channels)
-
-        for b, block in enumerate(self.control_model.down_blocks):
-            for r in range(len(block.resnets)):
-                increase_block_input_in_encoder_resnet(
-                    self.control_model, block_no=b, resnet_idx=r, by=next(it_extra_input_channels)
-                )
-
-            if block.downsamplers:
-                increase_block_input_in_encoder_downsampler(
-                    self.control_model, block_no=b, by=next(it_extra_input_channels)
-                )
-
-        increase_block_input_in_mid_resnet(self.control_model, by=extra_input_channels[-1])
-
-        # 2.3 - Make group norms work with modified channel sizes
-        adjust_group_norms(self.control_model)
-
-        # 3 - Gather Channel Sizes
-        self.ch_inout_ctrl = ControlNetXSModel._gather_subblock_sizes(self.control_model, base_or_control="control")
-        self.ch_inout_base = base_model_channel_sizes
-
-        # 4 - Build connections between base and control model
-        self.down_zero_convs_out = nn.ModuleList([])
-        self.down_zero_convs_in = nn.ModuleList([])
-        self.middle_block_out = nn.ModuleList([])
-        self.middle_block_in = nn.ModuleList([])
-        self.up_zero_convs_out = nn.ModuleList([])
-        self.up_zero_convs_in = nn.ModuleList([])
-
-        for ch_io_base in self.ch_inout_base["down"]:
-            self.down_zero_convs_in.append(self._make_zero_conv(in_channels=ch_io_base[1], out_channels=ch_io_base[1]))
-        for i in range(len(self.ch_inout_ctrl["down"])):
-            self.down_zero_convs_out.append(
-                self._make_zero_conv(self.ch_inout_ctrl["down"][i][1], self.ch_inout_base["down"][i][1])
-            )
-
-        self.middle_block_out = self._make_zero_conv(
-            self.ch_inout_ctrl["mid"][-1][1], self.ch_inout_base["mid"][-1][1]
-        )
-
-        self.up_zero_convs_out.append(
-            self._make_zero_conv(self.ch_inout_ctrl["down"][-1][1], self.ch_inout_base["mid"][-1][1])
-        )
-        for i in range(1, len(self.ch_inout_ctrl["down"])):
-            self.up_zero_convs_out.append(
-                self._make_zero_conv(self.ch_inout_ctrl["down"][-(i + 1)][1], self.ch_inout_base["up"][i - 1][1])
-            )
-
-        # 5 - Create conditioning hint embedding
-        self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
-            conditioning_embedding_channels=block_out_channels[0],
-            block_out_channels=conditioning_embedding_out_channels,
-            conditioning_channels=conditioning_channels,
-        )
-
-        # In the mininal implementation setting, we only need the control model up to the mid block
-        del self.control_model.up_blocks
-        del self.control_model.conv_norm_out
-        del self.control_model.conv_out
-
-    @classmethod
-    def from_unet(
-        cls,
-        unet: UNet2DConditionModel,
-        conditioning_channels: int = 3,
-        conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
-        controlnet_conditioning_channel_order: str = "rgb",
-        learn_embedding: bool = False,
-        time_embedding_mix: float = 1.0,
-        block_out_channels: Optional[Tuple[int]] = None,
-        size_ratio: Optional[float] = None,
-        num_attention_heads: Optional[Union[int, Tuple[int]]] = 8,
-        norm_num_groups: Optional[int] = None,
-    ):
-        r"""
-        Instantiate a [`ControlNetXSModel`] from [`UNet2DConditionModel`].
-
-        Parameters:
-            unet (`UNet2DConditionModel`):
-                The UNet model we want to control. The dimensions of the ControlNetXSModel will be adapted to it.
-            conditioning_channels (`int`, defaults to 3):
-                Number of channels of conditioning input (e.g. an image)
-            conditioning_embedding_out_channels (`tuple[int]`, defaults to `(16, 32, 96, 256)`):
-                The tuple of output channel for each block in the `controlnet_cond_embedding` layer.
-            controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
-                The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
-            learn_embedding (`bool`, defaults to `False`):
-                Wether to use time embedding of the control model. If yes, the time embedding is a linear interpolation
-                of the time embeddings of the control and base model with interpolation parameter
-                `time_embedding_mix**3`.
-            time_embedding_mix (`float`, defaults to 1.0):
-                Linear interpolation parameter used if `learn_embedding` is `True`.
-            block_out_channels (`Tuple[int]`, *optional*):
-                Down blocks output channels in control model. Either this or `size_ratio` must be given.
-            size_ratio (float, *optional*):
-                When given, block_out_channels is set to a relative fraction of the base model's block_out_channels.
-                Either this or `block_out_channels` must be given.
-            num_attention_heads (`Union[int, Tuple[int]]`, *optional*):
-                The dimension of the attention heads. The naming seems a bit confusing and it is, see https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 for why.
-            norm_num_groups (int, *optional*, defaults to `None`):
-                The number of groups to use for the normalization of the control unet. If `None`,
-                `int(unet.config.norm_num_groups * size_ratio)` is taken.
-        """
-
-        # Check input
-        fixed_size = block_out_channels is not None
-        relative_size = size_ratio is not None
-        if not (fixed_size ^ relative_size):
-            raise ValueError(
-                "Pass exactly one of `block_out_channels` (for absolute sizing) or `control_model_ratio` (for relative sizing)."
-            )
-
-        # Create model
-        if block_out_channels is None:
-            block_out_channels = [int(size_ratio * c) for c in unet.config.block_out_channels]
-
-        # Check that attention heads and group norms match channel sizes
-        # - attention heads
-        def attn_heads_match_channel_sizes(attn_heads, channel_sizes):
-            if isinstance(attn_heads, (tuple, list)):
-                return all(c % a == 0 for a, c in zip(attn_heads, channel_sizes))
-            else:
-                return all(c % attn_heads == 0 for c in channel_sizes)
-
-        num_attention_heads = num_attention_heads or unet.config.attention_head_dim
-        if not attn_heads_match_channel_sizes(num_attention_heads, block_out_channels):
-            raise ValueError(
-                f"The dimension of attention heads ({num_attention_heads}) must divide `block_out_channels` ({block_out_channels}). If you didn't set `num_attention_heads` the default settings don't match your model. Set `num_attention_heads` manually."
-            )
-
-        # - group norms
-        def group_norms_match_channel_sizes(num_groups, channel_sizes):
-            return all(c % num_groups == 0 for c in channel_sizes)
-
-        if norm_num_groups is None:
-            if group_norms_match_channel_sizes(unet.config.norm_num_groups, block_out_channels):
-                norm_num_groups = unet.config.norm_num_groups
-            else:
-                norm_num_groups = min(block_out_channels)
-
-                if group_norms_match_channel_sizes(norm_num_groups, block_out_channels):
-                    print(
-                        f"`norm_num_groups` was set to `min(block_out_channels)` (={norm_num_groups}) so it divides all block_out_channels` ({block_out_channels}). Set it explicitly to remove this information."
-                    )
-                else:
-                    raise ValueError(
-                        f"`block_out_channels` ({block_out_channels}) don't match the base models `norm_num_groups` ({unet.config.norm_num_groups}). Setting `norm_num_groups` to `min(block_out_channels)` ({norm_num_groups}) didn't fix this. Pass `norm_num_groups` explicitly so it divides all block_out_channels."
-                    )
-
-        def get_time_emb_input_dim(unet: UNet2DConditionModel):
-            return unet.time_embedding.linear_1.in_features
-
-        def get_time_emb_dim(unet: UNet2DConditionModel):
-            return unet.time_embedding.linear_2.out_features
-
-        # Clone params from base unet if
-        #    (i)   it's required to build SD or SDXL, and
-        #    (ii)  it's not used for the time embedding (as time embedding of control model is never used), and
-        #    (iii) it's not set further below anyway
-        to_keep = [
-            "cross_attention_dim",
-            "down_block_types",
-            "sample_size",
-            "transformer_layers_per_block",
-            "up_block_types",
-            "upcast_attention",
-        ]
-        kwargs = {k: v for k, v in dict(unet.config).items() if k in to_keep}
-        kwargs.update(block_out_channels=block_out_channels)
-        kwargs.update(num_attention_heads=num_attention_heads)
-        kwargs.update(norm_num_groups=norm_num_groups)
-
-        # Add controlnetxs-specific params
-        kwargs.update(
-            conditioning_channels=conditioning_channels,
-            controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
-            time_embedding_input_dim=get_time_emb_input_dim(unet),
-            time_embedding_dim=get_time_emb_dim(unet),
-            time_embedding_mix=time_embedding_mix,
-            learn_embedding=learn_embedding,
-            base_model_channel_sizes=ControlNetXSModel._gather_subblock_sizes(unet, base_or_control="base"),
-            conditioning_embedding_out_channels=conditioning_embedding_out_channels,
-        )
-
-        return cls(**kwargs)
-
-    @property
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        return self.control_model.attn_processors
-
-    def set_attn_processor(
-        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
-    ):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        self.control_model.set_attn_processor(processor, _remove_lora)
-
-    def set_default_attn_processor(self):
-        """
-        Disables custom attention processors and sets the default attention implementation.
-        """
-        self.control_model.set_default_attn_processor()
-
-    def set_attention_slice(self, slice_size):
-        r"""
-        Enable sliced attention computation.
-
-        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
-        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
-
-        Args:
-            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
-                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
-                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
-                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
-                must be a multiple of `slice_size`.
-        """
-        self.control_model.set_attention_slice(slice_size)
-
-    def _set_gradient_checkpointing(self, module, value=False):
-        if isinstance(module, (UNet2DConditionModel)):
-            if value:
-                module.enable_gradient_checkpointing()
-            else:
-                module.disable_gradient_checkpointing()
-
-    def forward(
-        self,
-        base_model: UNet2DConditionModel,
-        sample: torch.FloatTensor,
-        timestep: Union[torch.Tensor, float, int],
-        encoder_hidden_states: torch.Tensor,
-        controlnet_cond: torch.Tensor,
-        conditioning_scale: float = 1.0,
-        class_labels: Optional[torch.Tensor] = None,
-        timestep_cond: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
-        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
-        return_dict: bool = True,
-    ) -> Union[ControlNetXSOutput, Tuple]:
-        """
-        The [`ControlNetModel`] forward method.
-
-        Args:
-            base_model (`UNet2DConditionModel`):
-                The base unet model we want to control.
-            sample (`torch.FloatTensor`):
-                The noisy input tensor.
-            timestep (`Union[torch.Tensor, float, int]`):
-                The number of timesteps to denoise an input.
-            encoder_hidden_states (`torch.Tensor`):
-                The encoder hidden states.
-            controlnet_cond (`torch.FloatTensor`):
-                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
-            conditioning_scale (`float`, defaults to `1.0`):
-                How much the control model affects the base model outputs.
-            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
-                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
-            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
-                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
-                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
-                embeddings.
-            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
-                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
-                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
-                negative values to the attention scores corresponding to "discard" tokens.
-            added_cond_kwargs (`dict`):
-                Additional conditions for the Stable Diffusion XL UNet.
-            cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
-                A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
-            return_dict (`bool`, defaults to `True`):
-                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
-
-        Returns:
-            [`~models.controlnetxs.ControlNetXSOutput`] **or** `tuple`:
-                If `return_dict` is `True`, a [`~models.controlnetxs.ControlNetXSOutput`] is returned, otherwise a
-                tuple is returned where the first element is the sample tensor.
-        """
-        # check channel order
-        channel_order = self.config.controlnet_conditioning_channel_order
-
-        if channel_order == "rgb":
-            # in rgb order by default
-            ...
-        elif channel_order == "bgr":
-            controlnet_cond = torch.flip(controlnet_cond, dims=[1])
-        else:
-            raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
-
-        # scale control strength
-        n_connections = len(self.down_zero_convs_out) + 1 + len(self.up_zero_convs_out)
-        scale_list = torch.full((n_connections,), conditioning_scale)
-
-        # prepare attention_mask
-        if attention_mask is not None:
-            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
-            attention_mask = attention_mask.unsqueeze(1)
-
-        # 1. time
-        timesteps = timestep
-        if not torch.is_tensor(timesteps):
-            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
-            # This would be a good case for the `match` statement (Python 3.10+)
-            is_mps = sample.device.type == "mps"
-            if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
-            else:
-                dtype = torch.int32 if is_mps else torch.int64
-            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
-        elif len(timesteps.shape) == 0:
-            timesteps = timesteps[None].to(sample.device)
-
-        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
-        timesteps = timesteps.expand(sample.shape[0])
-
-        t_emb = base_model.time_proj(timesteps)
-
-        # timesteps does not contain any weights and will always return f32 tensors
-        # but time_embedding might actually be running in fp16. so we need to cast here.
-        # there might be better ways to encapsulate this.
-        t_emb = t_emb.to(dtype=sample.dtype)
-
-        if self.config.learn_embedding:
-            ctrl_temb = self.control_model.time_embedding(t_emb, timestep_cond)
-            base_temb = base_model.time_embedding(t_emb, timestep_cond)
-            interpolation_param = self.config.time_embedding_mix**0.3
-
-            temb = ctrl_temb * interpolation_param + base_temb * (1 - interpolation_param)
-        else:
-            temb = base_model.time_embedding(t_emb)
-
-        # added time & text embeddings
-        aug_emb = None
-
-        if base_model.class_embedding is not None:
-            if class_labels is None:
-                raise ValueError("class_labels should be provided when num_class_embeds > 0")
-
-            if base_model.config.class_embed_type == "timestep":
-                class_labels = base_model.time_proj(class_labels)
-
-            class_emb = base_model.class_embedding(class_labels).to(dtype=self.dtype)
-            temb = temb + class_emb
-
-        if base_model.config.addition_embed_type is not None:
-            if base_model.config.addition_embed_type == "text":
-                aug_emb = base_model.add_embedding(encoder_hidden_states)
-            elif base_model.config.addition_embed_type == "text_image":
-                raise NotImplementedError()
-            elif base_model.config.addition_embed_type == "text_time":
-                # SDXL - style
-                if "text_embeds" not in added_cond_kwargs:
-                    raise ValueError(
-                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
-                    )
-                text_embeds = added_cond_kwargs.get("text_embeds")
-                if "time_ids" not in added_cond_kwargs:
-                    raise ValueError(
-                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
-                    )
-                time_ids = added_cond_kwargs.get("time_ids")
-                time_embeds = base_model.add_time_proj(time_ids.flatten())
-                time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
-                add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
-                add_embeds = add_embeds.to(temb.dtype)
-                aug_emb = base_model.add_embedding(add_embeds)
-            elif base_model.config.addition_embed_type == "image":
-                raise NotImplementedError()
-            elif base_model.config.addition_embed_type == "image_hint":
-                raise NotImplementedError()
-
-        temb = temb + aug_emb if aug_emb is not None else temb
-
-        # text embeddings
-        cemb = encoder_hidden_states
-
-        # Preparation
-        guided_hint = self.controlnet_cond_embedding(controlnet_cond)
-
-        h_ctrl = h_base = sample
-        hs_base, hs_ctrl = [], []
-        it_down_convs_in, it_down_convs_out, it_dec_convs_in, it_up_convs_out = map(
-            iter, (self.down_zero_convs_in, self.down_zero_convs_out, self.up_zero_convs_in, self.up_zero_convs_out)
-        )
-        scales = iter(scale_list)
-
-        base_down_subblocks = to_sub_blocks(base_model.down_blocks)
-        ctrl_down_subblocks = to_sub_blocks(self.control_model.down_blocks)
-        base_mid_subblocks = to_sub_blocks([base_model.mid_block])
-        ctrl_mid_subblocks = to_sub_blocks([self.control_model.mid_block])
-        base_up_subblocks = to_sub_blocks(base_model.up_blocks)
-
-        # Cross Control
-        # 0 - conv in
-        h_base = base_model.conv_in(h_base)
-        h_ctrl = self.control_model.conv_in(h_ctrl)
-        if guided_hint is not None:
-            h_ctrl += guided_hint
-        h_base = h_base + next(it_down_convs_out)(h_ctrl) * next(scales)  # D - add ctrl -> base
-
-        hs_base.append(h_base)
-        hs_ctrl.append(h_ctrl)
-
-        # 1 - down
-        for m_base, m_ctrl in zip(base_down_subblocks, ctrl_down_subblocks):
-            h_ctrl = torch.cat([h_ctrl, next(it_down_convs_in)(h_base)], dim=1)  # A - concat base -> ctrl
-            h_base = m_base(h_base, temb, cemb, attention_mask, cross_attention_kwargs)  # B - apply base subblock
-            h_ctrl = m_ctrl(h_ctrl, temb, cemb, attention_mask, cross_attention_kwargs)  # C - apply ctrl subblock
-            h_base = h_base + next(it_down_convs_out)(h_ctrl) * next(scales)  # D - add ctrl -> base
-            hs_base.append(h_base)
-            hs_ctrl.append(h_ctrl)
-
-        # 2 - mid
-        h_ctrl = torch.cat([h_ctrl, next(it_down_convs_in)(h_base)], dim=1)  # A - concat base -> ctrl
-        for m_base, m_ctrl in zip(base_mid_subblocks, ctrl_mid_subblocks):
-            h_base = m_base(h_base, temb, cemb, attention_mask, cross_attention_kwargs)  # B - apply base subblock
-            h_ctrl = m_ctrl(h_ctrl, temb, cemb, attention_mask, cross_attention_kwargs)  # C - apply ctrl subblock
-        h_base = h_base + self.middle_block_out(h_ctrl) * next(scales)  # D - add ctrl -> base
-
-        # 3 - up
-        for i, m_base in enumerate(base_up_subblocks):
-            h_base = h_base + next(it_up_convs_out)(hs_ctrl.pop()) * next(scales)  # add info from ctrl encoder
-            h_base = torch.cat([h_base, hs_base.pop()], dim=1)  # concat info from base encoder+ctrl encoder
-            h_base = m_base(h_base, temb, cemb, attention_mask, cross_attention_kwargs)
-
-        h_base = base_model.conv_norm_out(h_base)
-        h_base = base_model.conv_act(h_base)
-        h_base = base_model.conv_out(h_base)
-
-        if not return_dict:
-            return h_base
-
-        return ControlNetXSOutput(sample=h_base)
-
-    def _make_zero_conv(self, in_channels, out_channels=None):
-        # keep running track of channels sizes
-        self.in_channels = in_channels
-        self.out_channels = out_channels or in_channels
-
-        return zero_module(nn.Conv2d(in_channels, out_channels, 1, padding=0))
-
-    @torch.no_grad()
-    def _check_if_vae_compatible(self, vae: AutoencoderKL):
-        condition_downscale_factor = 2 ** (len(self.config.conditioning_embedding_out_channels) - 1)
-        vae_downscale_factor = 2 ** (len(vae.config.block_out_channels) - 1)
-        compatible = condition_downscale_factor == vae_downscale_factor
-        return compatible, condition_downscale_factor, vae_downscale_factor
-
-
-class SubBlock(nn.ModuleList):
-    """A SubBlock is the largest piece of either base or control model, that is executed independently of the other model respectively.
-    Before each subblock, information is concatted from base to control. And after each subblock, information is added from control to base.
-    """
-
-    def __init__(self, ms, *args, **kwargs):
-        if not is_iterable(ms):
-            ms = [ms]
-        super().__init__(ms, *args, **kwargs)
-
-    def forward(
-        self,
-        x: torch.Tensor,
-        temb: torch.Tensor,
-        cemb: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
-    ):
-        """Iterate through children and pass correct information to each."""
-        for m in self:
-            if isinstance(m, ResnetBlock2D):
-                x = m(x, temb)
-            elif isinstance(m, Transformer2DModel):
-                x = m(x, cemb, attention_mask=attention_mask, cross_attention_kwargs=cross_attention_kwargs).sample
-            elif isinstance(m, Downsample2D):
-                x = m(x)
-            elif isinstance(m, Upsample2D):
-                x = m(x)
-            else:
-                raise ValueError(
-                    f"Type of m is {type(m)} but should be `ResnetBlock2D`, `Transformer2DModel`,  `Downsample2D` or `Upsample2D`"
-                )
-
-        return x
-
-
-def adjust_time_dims(unet: UNet2DConditionModel, in_dim: int, out_dim: int):
-    unet.time_embedding.linear_1 = nn.Linear(in_dim, out_dim)
-
-
-def increase_block_input_in_encoder_resnet(unet: UNet2DConditionModel, block_no, resnet_idx, by):
-    """Increase channels sizes to allow for additional concatted information from base model"""
-    r = unet.down_blocks[block_no].resnets[resnet_idx]
-    old_norm1, old_conv1 = r.norm1, r.conv1
-    # norm
-    norm_args = "num_groups num_channels eps affine".split(" ")
-    for a in norm_args:
-        assert hasattr(old_norm1, a)
-    norm_kwargs = {a: getattr(old_norm1, a) for a in norm_args}
-    norm_kwargs["num_channels"] += by  # surgery done here
-    # conv1
-    conv1_args = (
-        "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(" ")
-    )
-    for a in conv1_args:
-        assert hasattr(old_conv1, a)
-    conv1_kwargs = {a: getattr(old_conv1, a) for a in conv1_args}
-    conv1_kwargs["bias"] = "bias" in conv1_kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
-    conv1_kwargs["in_channels"] += by  # surgery done here
-    # conv_shortcut
-    # as we changed the input size of the block, the input and output sizes are likely different,
-    # therefore we need a conv_shortcut (simply adding won't work)
-    conv_shortcut_args_kwargs = {
-        "in_channels": conv1_kwargs["in_channels"],
-        "out_channels": conv1_kwargs["out_channels"],
-        # default arguments from resnet.__init__
-        "kernel_size": 1,
-        "stride": 1,
-        "padding": 0,
-        "bias": True,
-    }
-    # swap old with new modules
-    unet.down_blocks[block_no].resnets[resnet_idx].norm1 = GroupNorm(**norm_kwargs)
-    unet.down_blocks[block_no].resnets[resnet_idx].conv1 = LoRACompatibleConv(**conv1_kwargs)
-    unet.down_blocks[block_no].resnets[resnet_idx].conv_shortcut = LoRACompatibleConv(**conv_shortcut_args_kwargs)
-    unet.down_blocks[block_no].resnets[resnet_idx].in_channels += by  # surgery done here
-
-
-def increase_block_input_in_encoder_downsampler(unet: UNet2DConditionModel, block_no, by):
-    """Increase channels sizes to allow for additional concatted information from base model"""
-    old_down = unet.down_blocks[block_no].downsamplers[0].conv
-    # conv1
-    args = "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(
-        " "
-    )
-    for a in args:
-        assert hasattr(old_down, a)
-    kwargs = {a: getattr(old_down, a) for a in args}
-    kwargs["bias"] = "bias" in kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
-    kwargs["in_channels"] += by  # surgery done here
-    # swap old with new modules
-    unet.down_blocks[block_no].downsamplers[0].conv = LoRACompatibleConv(**kwargs)
-    unet.down_blocks[block_no].downsamplers[0].channels += by  # surgery done here
-
-
-def increase_block_input_in_mid_resnet(unet: UNet2DConditionModel, by):
-    """Increase channels sizes to allow for additional concatted information from base model"""
-    m = unet.mid_block.resnets[0]
-    old_norm1, old_conv1 = m.norm1, m.conv1
-    # norm
-    norm_args = "num_groups num_channels eps affine".split(" ")
-    for a in norm_args:
-        assert hasattr(old_norm1, a)
-    norm_kwargs = {a: getattr(old_norm1, a) for a in norm_args}
-    norm_kwargs["num_channels"] += by  # surgery done here
-    # conv1
-    conv1_args = (
-        "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(" ")
-    )
-    for a in conv1_args:
-        assert hasattr(old_conv1, a)
-    conv1_kwargs = {a: getattr(old_conv1, a) for a in conv1_args}
-    conv1_kwargs["bias"] = "bias" in conv1_kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
-    conv1_kwargs["in_channels"] += by  # surgery done here
-    # conv_shortcut
-    # as we changed the input size of the block, the input and output sizes are likely different,
-    # therefore we need a conv_shortcut (simply adding won't work)
-    conv_shortcut_args_kwargs = {
-        "in_channels": conv1_kwargs["in_channels"],
-        "out_channels": conv1_kwargs["out_channels"],
-        # default arguments from resnet.__init__
-        "kernel_size": 1,
-        "stride": 1,
-        "padding": 0,
-        "bias": True,
-    }
-    # swap old with new modules
-    unet.mid_block.resnets[0].norm1 = GroupNorm(**norm_kwargs)
-    unet.mid_block.resnets[0].conv1 = LoRACompatibleConv(**conv1_kwargs)
-    unet.mid_block.resnets[0].conv_shortcut = LoRACompatibleConv(**conv_shortcut_args_kwargs)
-    unet.mid_block.resnets[0].in_channels += by  # surgery done here
-
-
-def adjust_group_norms(unet: UNet2DConditionModel, max_num_group: int = 32):
-    def find_denominator(number, start):
-        if start >= number:
-            return number
-        while start != 0:
-            residual = number % start
-            if residual == 0:
-                return start
-            start -= 1
-
-    for block in [*unet.down_blocks, unet.mid_block]:
-        # resnets
-        for r in block.resnets:
-            if r.norm1.num_groups < max_num_group:
-                r.norm1.num_groups = find_denominator(r.norm1.num_channels, start=max_num_group)
-
-            if r.norm2.num_groups < max_num_group:
-                r.norm2.num_groups = find_denominator(r.norm2.num_channels, start=max_num_group)
-
-        # transformers
-        if hasattr(block, "attentions"):
-            for a in block.attentions:
-                if a.norm.num_groups < max_num_group:
-                    a.norm.num_groups = find_denominator(a.norm.num_channels, start=max_num_group)
-
-
-def is_iterable(o):
-    if isinstance(o, str):
-        return False
-    try:
-        iter(o)
-        return True
-    except TypeError:
-        return False
-
-
-def to_sub_blocks(blocks):
-    if not is_iterable(blocks):
-        blocks = [blocks]
-
-    sub_blocks = []
-
-    for b in blocks:
-        if hasattr(b, "resnets"):
-            if hasattr(b, "attentions") and b.attentions is not None:
-                for r, a in zip(b.resnets, b.attentions):
-                    sub_blocks.append([r, a])
-
-                num_resnets = len(b.resnets)
-                num_attns = len(b.attentions)
-
-                if num_resnets > num_attns:
-                    # we can have more resnets than attentions, so add each resnet as separate subblock
-                    for i in range(num_attns, num_resnets):
-                        sub_blocks.append([b.resnets[i]])
-            else:
-                for r in b.resnets:
-                    sub_blocks.append([r])
-
-        # upsamplers are part of the same subblock
-        if hasattr(b, "upsamplers") and b.upsamplers is not None:
-            for u in b.upsamplers:
-                sub_blocks[-1].extend([u])
-
-        # downsamplers are own subblock
-        if hasattr(b, "downsamplers") and b.downsamplers is not None:
-            for d in b.downsamplers:
-                sub_blocks.append([d])
-
-    return list(map(SubBlock, sub_blocks))
-
-
-def zero_module(module):
-    for p in module.parameters():
-        nn.init.zeros_(p)
-    return module

src/diffusers/models/embeddings.py

@@ -461,18 +461,6 @@ class ImageProjection(nn.Module):
         return image_embeds
 
 
-class MLPProjection(nn.Module):
-    def __init__(self, image_embed_dim=1024, cross_attention_dim=1024):
-        super().__init__()
-        from .attention import FeedForward
-
-        self.ff = FeedForward(image_embed_dim, cross_attention_dim, mult=1, activation_fn="gelu")
-        self.norm = nn.LayerNorm(cross_attention_dim)
-
-    def forward(self, image_embeds: torch.FloatTensor):
-        return self.norm(self.ff(image_embeds))
-
-
 class CombinedTimestepLabelEmbeddings(nn.Module):
     def __init__(self, num_classes, embedding_dim, class_dropout_prob=0.1):
         super().__init__()

src/diffusers/models/modeling_flax_utils.py

@@ -24,17 +24,13 @@ from flax.core.frozen_dict import FrozenDict, unfreeze
 from flax.serialization import from_bytes, to_bytes
 from flax.traverse_util import flatten_dict, unflatten_dict
 from huggingface_hub import create_repo, hf_hub_download
-from huggingface_hub.utils import (
-    EntryNotFoundError,
-    RepositoryNotFoundError,
-    RevisionNotFoundError,
-    validate_hf_hub_args,
-)
+from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
 from requests import HTTPError
 
 from .. import __version__, is_torch_available
 from ..utils import (
     CONFIG_NAME,
+    DIFFUSERS_CACHE,
     FLAX_WEIGHTS_NAME,
     HUGGINGFACE_CO_RESOLVE_ENDPOINT,
     WEIGHTS_NAME,
@@ -201,7 +197,6 @@ class FlaxModelMixin(PushToHubMixin):
         raise NotImplementedError(f"init_weights method has to be implemented for {self}")
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(
         cls,
         pretrained_model_name_or_path: Union[str, os.PathLike],
@@ -293,13 +288,13 @@ class FlaxModelMixin(PushToHubMixin):
         ```
         """
         config = kwargs.pop("config", None)
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         from_pt = kwargs.pop("from_pt", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         subfolder = kwargs.pop("subfolder", None)
 
@@ -319,7 +314,7 @@ class FlaxModelMixin(PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 subfolder=subfolder,
                 **kwargs,
@@ -364,7 +359,7 @@ class FlaxModelMixin(PushToHubMixin):
                     proxies=proxies,
                     resume_download=resume_download,
                     local_files_only=local_files_only,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     user_agent=user_agent,
                     subfolder=subfolder,
                     revision=revision,
@@ -374,7 +369,7 @@ class FlaxModelMixin(PushToHubMixin):
                 raise EnvironmentError(
                     f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier "
                     "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a "
-                    "token having permission to this repo with `token` or log in with `huggingface-cli "
+                    "token having permission to this repo with `use_auth_token` or log in with `huggingface-cli "
                     "login`."
                 )
             except RevisionNotFoundError:

src/diffusers/models/modeling_utils.py

@@ -25,13 +25,14 @@ from typing import Any, Callable, List, Optional, Tuple, Union
 import safetensors
 import torch
 from huggingface_hub import create_repo
-from huggingface_hub.utils import validate_hf_hub_args
 from torch import Tensor, nn
 
 from .. import __version__
 from ..utils import (
     CONFIG_NAME,
+    DIFFUSERS_CACHE,
     FLAX_WEIGHTS_NAME,
+    HF_HUB_OFFLINE,
     MIN_PEFT_VERSION,
     SAFETENSORS_WEIGHTS_NAME,
     WEIGHTS_NAME,
@@ -534,7 +535,6 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             )
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         r"""
         Instantiate a pretrained PyTorch model from a pretrained model configuration.
@@ -571,7 +571,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             local_files_only(`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -640,15 +640,15 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
         force_download = kwargs.pop("force_download", False)
         from_flax = kwargs.pop("from_flax", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         output_loading_info = kwargs.pop("output_loading_info", False)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
         subfolder = kwargs.pop("subfolder", None)
@@ -718,7 +718,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             resume_download=resume_download,
             proxies=proxies,
             local_files_only=local_files_only,
-            token=token,
+            use_auth_token=use_auth_token,
             revision=revision,
             subfolder=subfolder,
             device_map=device_map,
@@ -740,7 +740,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 subfolder=subfolder,
                 user_agent=user_agent,
@@ -763,7 +763,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                         resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
-                        token=token,
+                        use_auth_token=use_auth_token,
                         revision=revision,
                         subfolder=subfolder,
                         user_agent=user_agent,
@@ -782,7 +782,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                     resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     revision=revision,
                     subfolder=subfolder,
                     user_agent=user_agent,

src/diffusers/models/unet_2d_condition.py

@@ -25,7 +25,6 @@ from .activations import get_activation
 from .attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
-    Attention,
     AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
@@ -795,42 +794,6 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
                     setattr(upsample_block, k, None)
 
-    def fuse_qkv_projections(self):
-        """
-        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
-        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-        """
-        self.original_attn_processors = None
-
-        for _, attn_processor in self.attn_processors.items():
-            if "Added" in str(attn_processor.__class__.__name__):
-                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
-
-        self.original_attn_processors = self.attn_processors
-
-        for module in self.modules():
-            if isinstance(module, Attention):
-                module.fuse_projections(fuse=True)
-
-    def unfuse_qkv_projections(self):
-        """Disables the fused QKV projection if enabled.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        """
-        if self.original_attn_processors is not None:
-            self.set_attn_processor(self.original_attn_processors)
-
     def forward(
         self,
         sample: torch.FloatTensor,

src/diffusers/pipelines/__init__.py

@@ -19,7 +19,6 @@ from ..utils import (
 _dummy_objects = {}
 _import_structure = {
     "controlnet": [],
-    "controlnet_xs": [],
     "latent_diffusion": [],
     "stable_diffusion": [],
     "stable_diffusion_xl": [],
@@ -94,12 +93,6 @@ else:
             "StableDiffusionXLControlNetPipeline",
         ]
     )
-    _import_structure["controlnet_xs"].extend(
-        [
-            "StableDiffusionControlNetXSPipeline",
-            "StableDiffusionXLControlNetXSPipeline",
-        ]
-    )
     _import_structure["deepfloyd_if"] = [
         "IFImg2ImgPipeline",
         "IFImg2ImgSuperResolutionPipeline",
@@ -354,10 +347,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionXLControlNetInpaintPipeline,
             StableDiffusionXLControlNetPipeline,
         )
-        from .controlnet_xs import (
-            StableDiffusionControlNetXSPipeline,
-            StableDiffusionXLControlNetXSPipeline,
-        )
         from .deepfloyd_if import (
             IFImg2ImgPipeline,
             IFImg2ImgSuperResolutionPipeline,

src/diffusers/pipelines/auto_pipeline.py

@@ -16,9 +16,8 @@
 import inspect
 from collections import OrderedDict
 
-from huggingface_hub.utils import validate_hf_hub_args
-
 from ..configuration_utils import ConfigMixin
+from ..utils import DIFFUSERS_CACHE
 from .controlnet import (
     StableDiffusionControlNetImg2ImgPipeline,
     StableDiffusionControlNetInpaintPipeline,
@@ -196,7 +195,6 @@ class AutoPipelineForText2Image(ConfigMixin):
         )
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_or_path, **kwargs):
         r"""
         Instantiates a text-to-image Pytorch diffusion pipeline from pretrained pipeline weight.
@@ -248,7 +246,7 @@ class AutoPipelineForText2Image(ConfigMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -312,11 +310,11 @@ class AutoPipelineForText2Image(ConfigMixin):
         >>> image = pipeline(prompt).images[0]
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
 
@@ -325,7 +323,7 @@ class AutoPipelineForText2Image(ConfigMixin):
             "force_download": force_download,
             "resume_download": resume_download,
             "proxies": proxies,
-            "token": token,
+            "use_auth_token": use_auth_token,
             "local_files_only": local_files_only,
             "revision": revision,
         }
@@ -468,7 +466,6 @@ class AutoPipelineForImage2Image(ConfigMixin):
         )
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_or_path, **kwargs):
         r"""
         Instantiates a image-to-image Pytorch diffusion pipeline from pretrained pipeline weight.
@@ -521,7 +518,7 @@ class AutoPipelineForImage2Image(ConfigMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -585,11 +582,11 @@ class AutoPipelineForImage2Image(ConfigMixin):
         >>> image = pipeline(prompt, image).images[0]
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
 
@@ -598,7 +595,7 @@ class AutoPipelineForImage2Image(ConfigMixin):
             "force_download": force_download,
             "resume_download": resume_download,
             "proxies": proxies,
-            "token": token,
+            "use_auth_token": use_auth_token,
             "local_files_only": local_files_only,
             "revision": revision,
         }
@@ -745,7 +742,6 @@ class AutoPipelineForInpainting(ConfigMixin):
         )
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_or_path, **kwargs):
         r"""
         Instantiates a inpainting Pytorch diffusion pipeline from pretrained pipeline weight.
@@ -797,7 +793,7 @@ class AutoPipelineForInpainting(ConfigMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -861,11 +857,11 @@ class AutoPipelineForInpainting(ConfigMixin):
         >>> image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0]
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         force_download = kwargs.pop("force_download", False)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
 
@@ -874,7 +870,7 @@ class AutoPipelineForInpainting(ConfigMixin):
             "force_download": force_download,
             "resume_download": resume_download,
             "proxies": proxies,
-            "token": token,
+            "use_auth_token": use_auth_token,
             "local_files_only": local_files_only,
             "revision": revision,
         }

src/diffusers/pipelines/controlnet_xs/__init__.py

@@ -1,68 +0,0 @@
-from typing import TYPE_CHECKING
-
-from ...utils import (
-    DIFFUSERS_SLOW_IMPORT,
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    get_objects_from_module,
-    is_flax_available,
-    is_torch_available,
-    is_transformers_available,
-)
-
-
-_dummy_objects = {}
-_import_structure = {}
-
-try:
-    if not (is_transformers_available() and is_torch_available()):
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
-
-    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
-else:
-    _import_structure["pipeline_controlnet_xs"] = ["StableDiffusionControlNetXSPipeline"]
-    _import_structure["pipeline_controlnet_xs_sd_xl"] = ["StableDiffusionXLControlNetXSPipeline"]
-try:
-    if not (is_transformers_available() and is_flax_available()):
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    from ...utils import dummy_flax_and_transformers_objects  # noqa F403
-
-    _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects))
-else:
-    pass  # _import_structure["pipeline_flax_controlnet"] = ["FlaxStableDiffusionControlNetPipeline"]
-
-
-if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
-    try:
-        if not (is_transformers_available() and is_torch_available()):
-            raise OptionalDependencyNotAvailable()
-
-    except OptionalDependencyNotAvailable:
-        from ...utils.dummy_torch_and_transformers_objects import *
-    else:
-        from .pipeline_controlnet_xs import StableDiffusionControlNetXSPipeline
-        from .pipeline_controlnet_xs_sd_xl import StableDiffusionXLControlNetXSPipeline
-
-    try:
-        if not (is_transformers_available() and is_flax_available()):
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        from ...utils.dummy_flax_and_transformers_objects import *  # noqa F403
-    else:
-        pass  # from .pipeline_flax_controlnet import FlaxStableDiffusionControlNetPipeline
-
-
-else:
-    import sys
-
-    sys.modules[__name__] = _LazyModule(
-        __name__,
-        globals()["__file__"],
-        _import_structure,
-        module_spec=__spec__,
-    )
-    for name, value in _dummy_objects.items():
-        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs.py

@@ -1,944 +0,0 @@
-# Copyright 2023 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import inspect
-from typing import Any, Callable, Dict, List, Optional, Union
-
-import numpy as np
-import PIL.Image
-import torch
-import torch.nn.functional as F
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
-
-from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
-from ...models import AutoencoderKL, ControlNetXSModel, UNet2DConditionModel
-from ...models.lora import adjust_lora_scale_text_encoder
-from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import (
-    USE_PEFT_BACKEND,
-    deprecate,
-    logging,
-    replace_example_docstring,
-    scale_lora_layers,
-    unscale_lora_layers,
-)
-from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
-from ..pipeline_utils import DiffusionPipeline
-from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
-from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
-
-
-logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
-
-
-EXAMPLE_DOC_STRING = """
-    Examples:
-        ```py
-        >>> # !pip install opencv-python transformers accelerate
-        >>> from diffusers import StableDiffusionControlNetXSPipeline, ControlNetXSModel
-        >>> from diffusers.utils import load_image
-        >>> import numpy as np
-        >>> import torch
-
-        >>> import cv2
-        >>> from PIL import Image
-
-        >>> prompt = "aerial view, a futuristic research complex in a bright foggy jungle, hard lighting"
-        >>> negative_prompt = "low quality, bad quality, sketches"
-
-        >>> # download an image
-        >>> image = load_image(
-        ...    "https://hf.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png"
-        ... )
-
-        >>> # initialize the models and pipeline
-        >>> controlnet_conditioning_scale = 0.5
-        >>> controlnet = ControlNetXSModel.from_pretrained(
-        ...     "UmerHA/ConrolNetXS-SD2.1-canny", torch_dtype=torch.float16
-        ... )
-        >>> pipe = StableDiffusionControlNetXSPipeline.from_pretrained(
-        ...     "stabilityai/stable-diffusion-2-1", controlnet=controlnet, torch_dtype=torch.float16
-        ... )
-        >>> pipe.enable_model_cpu_offload()
-
-        >>> # get canny image
-        >>> image = np.array(image)
-        >>> image = cv2.Canny(image, 100, 200)
-        >>> image = image[:, :, None]
-        >>> image = np.concatenate([image, image, image], axis=2)
-        >>> canny_image = Image.fromarray(image)
-        >>> # generate image
-        >>> image = pipe(
-        ...     prompt, controlnet_conditioning_scale=controlnet_conditioning_scale, image=canny_image
-        ... ).images[0]
-        ```
-"""
-
-
-class StableDiffusionControlNetXSPipeline(
-    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
-):
-    r"""
-    Pipeline for text-to-image generation using Stable Diffusion with ControlNet-XS guidance.
-
-    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
-    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
-
-    The pipeline also inherits the following loading methods:
-        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
-        - [`loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
-
-    Args:
-        vae ([`AutoencoderKL`]):
-            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
-        text_encoder ([`~transformers.CLIPTextModel`]):
-            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
-        tokenizer ([`~transformers.CLIPTokenizer`]):
-            A `CLIPTokenizer` to tokenize text.
-        unet ([`UNet2DConditionModel`]):
-            A `UNet2DConditionModel` to denoise the encoded image latents.
-        controlnet ([`ControlNetXSModel`]):
-            Provides additional conditioning to the `unet` during the denoising process.
-        scheduler ([`SchedulerMixin`]):
-            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
-            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
-        safety_checker ([`StableDiffusionSafetyChecker`]):
-            Classification module that estimates whether generated images could be considered offensive or harmful.
-            Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
-            about a model's potential harms.
-        feature_extractor ([`~transformers.CLIPImageProcessor`]):
-            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
-    """
-
-    model_cpu_offload_seq = "text_encoder->unet->vae>controlnet"
-    _optional_components = ["safety_checker", "feature_extractor"]
-    _exclude_from_cpu_offload = ["safety_checker"]
-
-    def __init__(
-        self,
-        vae: AutoencoderKL,
-        text_encoder: CLIPTextModel,
-        tokenizer: CLIPTokenizer,
-        unet: UNet2DConditionModel,
-        controlnet: ControlNetXSModel,
-        scheduler: KarrasDiffusionSchedulers,
-        safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPImageProcessor,
-        requires_safety_checker: bool = True,
-    ):
-        super().__init__()
-
-        if safety_checker is None and requires_safety_checker:
-            logger.warning(
-                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
-                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
-                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
-                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
-                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
-                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
-            )
-
-        if safety_checker is not None and feature_extractor is None:
-            raise ValueError(
-                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
-                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
-            )
-
-        vae_compatible, cnxs_condition_downsample_factor, vae_downsample_factor = controlnet._check_if_vae_compatible(
-            vae
-        )
-        if not vae_compatible:
-            raise ValueError(
-                f"The downsampling factors of the VAE ({vae_downsample_factor}) and the conditioning part of ControlNetXS model {cnxs_condition_downsample_factor} need to be equal. Consider building the ControlNetXS model with different `conditioning_block_sizes`."
-            )
-
-        self.register_modules(
-            vae=vae,
-            text_encoder=text_encoder,
-            tokenizer=tokenizer,
-            unet=unet,
-            controlnet=controlnet,
-            scheduler=scheduler,
-            safety_checker=safety_checker,
-            feature_extractor=feature_extractor,
-        )
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
-        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
-        self.control_image_processor = VaeImageProcessor(
-            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
-        )
-        self.register_to_config(requires_safety_checker=requires_safety_checker)
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
-    def enable_vae_slicing(self):
-        r"""
-        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
-        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
-        """
-        self.vae.enable_slicing()
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
-    def disable_vae_slicing(self):
-        r"""
-        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
-        computing decoding in one step.
-        """
-        self.vae.disable_slicing()
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
-    def enable_vae_tiling(self):
-        r"""
-        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
-        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
-        processing larger images.
-        """
-        self.vae.enable_tiling()
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
-    def disable_vae_tiling(self):
-        r"""
-        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
-        computing decoding in one step.
-        """
-        self.vae.disable_tiling()
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
-    def _encode_prompt(
-        self,
-        prompt,
-        device,
-        num_images_per_prompt,
-        do_classifier_free_guidance,
-        negative_prompt=None,
-        prompt_embeds: Optional[torch.FloatTensor] = None,
-        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
-        lora_scale: Optional[float] = None,
-        **kwargs,
-    ):
-        deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
-        deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
-
-        prompt_embeds_tuple = self.encode_prompt(
-            prompt=prompt,
-            device=device,
-            num_images_per_prompt=num_images_per_prompt,
-            do_classifier_free_guidance=do_classifier_free_guidance,
-            negative_prompt=negative_prompt,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            lora_scale=lora_scale,
-            **kwargs,
-        )
-
-        # concatenate for backwards comp
-        prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
-
-        return prompt_embeds
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
-    def encode_prompt(
-        self,
-        prompt,
-        device,
-        num_images_per_prompt,
-        do_classifier_free_guidance,
-        negative_prompt=None,
-        prompt_embeds: Optional[torch.FloatTensor] = None,
-        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
-        lora_scale: Optional[float] = None,
-        clip_skip: Optional[int] = None,
-    ):
-        r"""
-        Encodes the prompt into text encoder hidden states.
-
-        Args:
-            prompt (`str` or `List[str]`, *optional*):
-                prompt to be encoded
-            device: (`torch.device`):
-                torch device
-            num_images_per_prompt (`int`):
-                number of images that should be generated per prompt
-            do_classifier_free_guidance (`bool`):
-                whether to use classifier free guidance or not
-            negative_prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts not to guide the image generation. If not defined, one has to pass
-                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
-                less than `1`).
-            prompt_embeds (`torch.FloatTensor`, *optional*):
-                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
-                provided, text embeddings will be generated from `prompt` input argument.
-            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
-                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
-                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
-                argument.
-            lora_scale (`float`, *optional*):
-                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
-            clip_skip (`int`, *optional*):
-                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
-                the output of the pre-final layer will be used for computing the prompt embeddings.
-        """
-        # set lora scale so that monkey patched LoRA
-        # function of text encoder can correctly access it
-        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
-            self._lora_scale = lora_scale
-
-            # dynamically adjust the LoRA scale
-            if not USE_PEFT_BACKEND:
-                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
-            else:
-                scale_lora_layers(self.text_encoder, lora_scale)
-
-        if prompt is not None and isinstance(prompt, str):
-            batch_size = 1
-        elif prompt is not None and isinstance(prompt, list):
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
-
-        if prompt_embeds is None:
-            # textual inversion: procecss multi-vector tokens if necessary
-            if isinstance(self, TextualInversionLoaderMixin):
-                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
-
-            text_inputs = self.tokenizer(
-                prompt,
-                padding="max_length",
-                max_length=self.tokenizer.model_max_length,
-                truncation=True,
-                return_tensors="pt",
-            )
-            text_input_ids = text_inputs.input_ids
-            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
-
-            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
-                text_input_ids, untruncated_ids
-            ):
-                removed_text = self.tokenizer.batch_decode(
-                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
-                )
-                logger.warning(
-                    "The following part of your input was truncated because CLIP can only handle sequences up to"
-                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
-                )
-
-            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
-                attention_mask = text_inputs.attention_mask.to(device)
-            else:
-                attention_mask = None
-
-            if clip_skip is None:
-                prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
-                prompt_embeds = prompt_embeds[0]
-            else:
-                prompt_embeds = self.text_encoder(
-                    text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
-                )
-                # Access the `hidden_states` first, that contains a tuple of
-                # all the hidden states from the encoder layers. Then index into
-                # the tuple to access the hidden states from the desired layer.
-                prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
-                # We also need to apply the final LayerNorm here to not mess with the
-                # representations. The `last_hidden_states` that we typically use for
-                # obtaining the final prompt representations passes through the LayerNorm
-                # layer.
-                prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
-
-        if self.text_encoder is not None:
-            prompt_embeds_dtype = self.text_encoder.dtype
-        elif self.unet is not None:
-            prompt_embeds_dtype = self.unet.dtype
-        else:
-            prompt_embeds_dtype = prompt_embeds.dtype
-
-        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
-
-        bs_embed, seq_len, _ = prompt_embeds.shape
-        # duplicate text embeddings for each generation per prompt, using mps friendly method
-        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
-        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
-
-        # get unconditional embeddings for classifier free guidance
-        if do_classifier_free_guidance and negative_prompt_embeds is None:
-            uncond_tokens: List[str]
-            if negative_prompt is None:
-                uncond_tokens = [""] * batch_size
-            elif prompt is not None and type(prompt) is not type(negative_prompt):
-                raise TypeError(
-                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
-                    f" {type(prompt)}."
-                )
-            elif isinstance(negative_prompt, str):
-                uncond_tokens = [negative_prompt]
-            elif batch_size != len(negative_prompt):
-                raise ValueError(
-                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
-                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
-                    " the batch size of `prompt`."
-                )
-            else:
-                uncond_tokens = negative_prompt
-
-            # textual inversion: procecss multi-vector tokens if necessary
-            if isinstance(self, TextualInversionLoaderMixin):
-                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
-
-            max_length = prompt_embeds.shape[1]
-            uncond_input = self.tokenizer(
-                uncond_tokens,
-                padding="max_length",
-                max_length=max_length,
-                truncation=True,
-                return_tensors="pt",
-            )
-
-            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
-                attention_mask = uncond_input.attention_mask.to(device)
-            else:
-                attention_mask = None
-
-            negative_prompt_embeds = self.text_encoder(
-                uncond_input.input_ids.to(device),
-                attention_mask=attention_mask,
-            )
-            negative_prompt_embeds = negative_prompt_embeds[0]
-
-        if do_classifier_free_guidance:
-            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
-            seq_len = negative_prompt_embeds.shape[1]
-
-            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
-
-            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
-            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
-
-        if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
-            # Retrieve the original scale by scaling back the LoRA layers
-            unscale_lora_layers(self.text_encoder, lora_scale)
-
-        return prompt_embeds, negative_prompt_embeds
-
-    # 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:
-            has_nsfw_concept = None
-        else:
-            if torch.is_tensor(image):
-                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
-            else:
-                feature_extractor_input = self.image_processor.numpy_to_pil(image)
-            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
-            image, has_nsfw_concept = self.safety_checker(
-                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
-            )
-        return image, has_nsfw_concept
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
-    def decode_latents(self, latents):
-        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
-        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
-
-        latents = 1 / self.vae.config.scaling_factor * latents
-        image = self.vae.decode(latents, return_dict=False)[0]
-        image = (image / 2 + 0.5).clamp(0, 1)
-        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
-        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
-        return image
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
-    def prepare_extra_step_kwargs(self, generator, eta):
-        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
-        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
-        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
-        # and should be between [0, 1]
-
-        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
-        extra_step_kwargs = {}
-        if accepts_eta:
-            extra_step_kwargs["eta"] = eta
-
-        # check if the scheduler accepts generator
-        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
-        if accepts_generator:
-            extra_step_kwargs["generator"] = generator
-        return extra_step_kwargs
-
-    def check_inputs(
-        self,
-        prompt,
-        image,
-        callback_steps,
-        negative_prompt=None,
-        prompt_embeds=None,
-        negative_prompt_embeds=None,
-        controlnet_conditioning_scale=1.0,
-        control_guidance_start=0.0,
-        control_guidance_end=1.0,
-    ):
-        if (callback_steps is None) or (
-            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
-        ):
-            raise ValueError(
-                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
-                f" {type(callback_steps)}."
-            )
-
-        if prompt is not None and prompt_embeds is not None:
-            raise ValueError(
-                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
-                " only forward one of the two."
-            )
-        elif prompt is None and prompt_embeds is None:
-            raise ValueError(
-                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
-            )
-        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
-            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
-
-        if negative_prompt is not None and negative_prompt_embeds is not None:
-            raise ValueError(
-                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
-                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
-            )
-
-        if prompt_embeds is not None and negative_prompt_embeds is not None:
-            if prompt_embeds.shape != negative_prompt_embeds.shape:
-                raise ValueError(
-                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
-                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
-                    f" {negative_prompt_embeds.shape}."
-                )
-
-        # Check `image`
-        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
-            self.controlnet, torch._dynamo.eval_frame.OptimizedModule
-        )
-        if (
-            isinstance(self.controlnet, ControlNetXSModel)
-            or is_compiled
-            and isinstance(self.controlnet._orig_mod, ControlNetXSModel)
-        ):
-            self.check_image(image, prompt, prompt_embeds)
-        else:
-            assert False
-
-        # Check `controlnet_conditioning_scale`
-        if (
-            isinstance(self.controlnet, ControlNetXSModel)
-            or is_compiled
-            and isinstance(self.controlnet._orig_mod, ControlNetXSModel)
-        ):
-            if not isinstance(controlnet_conditioning_scale, float):
-                raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
-        else:
-            assert False
-
-        start, end = control_guidance_start, control_guidance_end
-        if start >= end:
-            raise ValueError(
-                f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
-            )
-        if start < 0.0:
-            raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
-        if end > 1.0:
-            raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
-
-    def check_image(self, image, prompt, prompt_embeds):
-        image_is_pil = isinstance(image, PIL.Image.Image)
-        image_is_tensor = isinstance(image, torch.Tensor)
-        image_is_np = isinstance(image, np.ndarray)
-        image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
-        image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
-        image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
-
-        if (
-            not image_is_pil
-            and not image_is_tensor
-            and not image_is_np
-            and not image_is_pil_list
-            and not image_is_tensor_list
-            and not image_is_np_list
-        ):
-            raise TypeError(
-                f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
-            )
-
-        if image_is_pil:
-            image_batch_size = 1
-        else:
-            image_batch_size = len(image)
-
-        if prompt is not None and isinstance(prompt, str):
-            prompt_batch_size = 1
-        elif prompt is not None and isinstance(prompt, list):
-            prompt_batch_size = len(prompt)
-        elif prompt_embeds is not None:
-            prompt_batch_size = prompt_embeds.shape[0]
-
-        if image_batch_size != 1 and image_batch_size != prompt_batch_size:
-            raise ValueError(
-                f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
-            )
-
-    def prepare_image(
-        self,
-        image,
-        width,
-        height,
-        batch_size,
-        num_images_per_prompt,
-        device,
-        dtype,
-        do_classifier_free_guidance=False,
-    ):
-        image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
-        image_batch_size = image.shape[0]
-
-        if image_batch_size == 1:
-            repeat_by = batch_size
-        else:
-            # image batch size is the same as prompt batch size
-            repeat_by = num_images_per_prompt
-
-        image = image.repeat_interleave(repeat_by, dim=0)
-
-        image = image.to(device=device, dtype=dtype)
-
-        if do_classifier_free_guidance:
-            image = torch.cat([image] * 2)
-
-        return image
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
-    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
-            )
-
-        if latents is None:
-            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
-        else:
-            latents = latents.to(device)
-
-        # scale the initial noise by the standard deviation required by the scheduler
-        latents = latents * self.scheduler.init_noise_sigma
-        return latents
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
-    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
-        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
-
-        The suffixes after the scaling factors represent the stages where they are being applied.
-
-        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
-        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
-
-        Args:
-            s1 (`float`):
-                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
-                mitigate "oversmoothing effect" in the enhanced denoising process.
-            s2 (`float`):
-                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
-                mitigate "oversmoothing effect" in the enhanced denoising process.
-            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
-            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
-        """
-        if not hasattr(self, "unet"):
-            raise ValueError("The pipeline must have `unet` for using FreeU.")
-        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
-    def disable_freeu(self):
-        """Disables the FreeU mechanism if enabled."""
-        self.unet.disable_freeu()
-
-    @torch.no_grad()
-    @replace_example_docstring(EXAMPLE_DOC_STRING)
-    def __call__(
-        self,
-        prompt: Union[str, List[str]] = None,
-        image: PipelineImageInput = None,
-        height: Optional[int] = None,
-        width: Optional[int] = None,
-        num_inference_steps: int = 50,
-        guidance_scale: float = 7.5,
-        negative_prompt: Optional[Union[str, List[str]]] = None,
-        num_images_per_prompt: Optional[int] = 1,
-        eta: float = 0.0,
-        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
-        latents: Optional[torch.FloatTensor] = None,
-        prompt_embeds: Optional[torch.FloatTensor] = None,
-        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
-        output_type: Optional[str] = "pil",
-        return_dict: bool = True,
-        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
-        callback_steps: int = 1,
-        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
-        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
-        control_guidance_start: float = 0.0,
-        control_guidance_end: float = 1.0,
-        clip_skip: Optional[int] = None,
-    ):
-        r"""
-        The call function to the pipeline for generation.
-
-        Args:
-            prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
-            image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,
-                    `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
-                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
-                specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
-                accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
-                and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
-                `init`, images must be passed as a list such that each element of the list can be correctly batched for
-                input to a single ControlNet.
-            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
-                The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
-                The width in pixels of the generated image.
-            num_inference_steps (`int`, *optional*, defaults to 50):
-                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
-                expense of slower inference.
-            guidance_scale (`float`, *optional*, defaults to 7.5):
-                A higher guidance scale value encourages the model to generate images closely linked to the text
-                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
-            negative_prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
-                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
-            num_images_per_prompt (`int`, *optional*, defaults to 1):
-                The number of images to generate per prompt.
-            eta (`float`, *optional*, defaults to 0.0):
-                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
-                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
-            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
-                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
-                generation deterministic.
-            latents (`torch.FloatTensor`, *optional*):
-                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
-                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
-                tensor is generated by sampling using the supplied random `generator`.
-            prompt_embeds (`torch.FloatTensor`, *optional*):
-                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
-                provided, text embeddings are generated from the `prompt` input argument.
-            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
-                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
-                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
-            output_type (`str`, *optional*, defaults to `"pil"`):
-                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
-            return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
-                plain tuple.
-            callback (`Callable`, *optional*):
-                A function that calls every `callback_steps` steps during inference. The function is called with the
-                following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
-            callback_steps (`int`, *optional*, defaults to 1):
-                The frequency at which the `callback` function is called. If not specified, the callback is called at
-                every step.
-            cross_attention_kwargs (`dict`, *optional*):
-                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
-                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
-            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
-                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
-                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
-                the corresponding scale as a list.
-            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
-                The percentage of total steps at which the ControlNet starts applying.
-            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
-                The percentage of total steps at which the ControlNet stops applying.
-            clip_skip (`int`, *optional*):
-                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
-                the output of the pre-final layer will be used for computing the prompt embeddings.
-
-        Examples:
-
-        Returns:
-            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
-                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
-                otherwise a `tuple` is returned where the first element is a list with the generated images and the
-                second element is a list of `bool`s indicating whether the corresponding generated image contains
-                "not-safe-for-work" (nsfw) content.
-        """
-        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
-
-        # 1. Check inputs. Raise error if not correct
-        self.check_inputs(
-            prompt,
-            image,
-            callback_steps,
-            negative_prompt,
-            prompt_embeds,
-            negative_prompt_embeds,
-            controlnet_conditioning_scale,
-            control_guidance_start,
-            control_guidance_end,
-        )
-
-        # 2. Define call parameters
-        if prompt is not None and isinstance(prompt, str):
-            batch_size = 1
-        elif prompt is not None and isinstance(prompt, list):
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
-
-        device = self._execution_device
-        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
-        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
-        # corresponds to doing no classifier free guidance.
-        do_classifier_free_guidance = guidance_scale > 1.0
-
-        # 3. Encode input prompt
-        text_encoder_lora_scale = (
-            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
-        )
-        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
-            prompt,
-            device,
-            num_images_per_prompt,
-            do_classifier_free_guidance,
-            negative_prompt,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            lora_scale=text_encoder_lora_scale,
-            clip_skip=clip_skip,
-        )
-        # For classifier free guidance, we need to do two forward passes.
-        # Here we concatenate the unconditional and text embeddings into a single batch
-        # to avoid doing two forward passes
-        if do_classifier_free_guidance:
-            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
-
-        # 4. Prepare image
-        if isinstance(controlnet, ControlNetXSModel):
-            image = self.prepare_image(
-                image=image,
-                width=width,
-                height=height,
-                batch_size=batch_size * num_images_per_prompt,
-                num_images_per_prompt=num_images_per_prompt,
-                device=device,
-                dtype=controlnet.dtype,
-                do_classifier_free_guidance=do_classifier_free_guidance,
-            )
-            height, width = image.shape[-2:]
-        else:
-            assert False
-
-        # 5. Prepare timesteps
-        self.scheduler.set_timesteps(num_inference_steps, device=device)
-        timesteps = self.scheduler.timesteps
-
-        # 6. Prepare latent variables
-        num_channels_latents = self.unet.config.in_channels
-        latents = self.prepare_latents(
-            batch_size * num_images_per_prompt,
-            num_channels_latents,
-            height,
-            width,
-            prompt_embeds.dtype,
-            device,
-            generator,
-            latents,
-        )
-
-        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
-        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
-
-        # 8. Denoising loop
-        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
-        is_unet_compiled = is_compiled_module(self.unet)
-        is_controlnet_compiled = is_compiled_module(self.controlnet)
-        is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
-        with self.progress_bar(total=num_inference_steps) as progress_bar:
-            for i, t in enumerate(timesteps):
-                # Relevant thread:
-                # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
-                if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
-                    torch._inductor.cudagraph_mark_step_begin()
-                # expand the latents if we are doing classifier free guidance
-                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
-                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-
-                # predict the noise residual
-                dont_control = (
-                    i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end
-                )
-                if dont_control:
-                    noise_pred = self.unet(
-                        sample=latent_model_input,
-                        timestep=t,
-                        encoder_hidden_states=prompt_embeds,
-                        cross_attention_kwargs=cross_attention_kwargs,
-                        return_dict=True,
-                    ).sample
-                else:
-                    noise_pred = self.controlnet(
-                        base_model=self.unet,
-                        sample=latent_model_input,
-                        timestep=t,
-                        encoder_hidden_states=prompt_embeds,
-                        controlnet_cond=image,
-                        conditioning_scale=controlnet_conditioning_scale,
-                        cross_attention_kwargs=cross_attention_kwargs,
-                        return_dict=True,
-                    ).sample
-
-                # perform guidance
-                if do_classifier_free_guidance:
-                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
-
-                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
-
-                # call the callback, if provided
-                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
-                    progress_bar.update()
-                    if callback is not None and i % callback_steps == 0:
-                        step_idx = i // getattr(self.scheduler, "order", 1)
-                        callback(step_idx, t, latents)
-
-        # If we do sequential model offloading, let's offload unet and controlnet
-        # manually for max memory savings
-        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
-            self.unet.to("cpu")
-            self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
-
-        if not output_type == "latent":
-            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
-                0
-            ]
-            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
-        else:
-            image = latents
-            has_nsfw_concept = None
-
-        if has_nsfw_concept is None:
-            do_denormalize = [True] * image.shape[0]
-        else:
-            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
-
-        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
-
-        # Offload all models
-        self.maybe_free_model_hooks()
-
-        if not return_dict:
-            return (image, has_nsfw_concept)
-
-        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

src/diffusers/pipelines/onnx_utils.py

@@ -22,7 +22,6 @@ from typing import Optional, Union
 
 import numpy as np
 from huggingface_hub import hf_hub_download
-from huggingface_hub.utils import validate_hf_hub_args
 
 from ..utils import ONNX_EXTERNAL_WEIGHTS_NAME, ONNX_WEIGHTS_NAME, is_onnx_available, logging
 
@@ -131,11 +130,10 @@ class OnnxRuntimeModel:
         self._save_pretrained(save_directory, **kwargs)
 
     @classmethod
-    @validate_hf_hub_args
     def _from_pretrained(
         cls,
         model_id: Union[str, Path],
-        token: Optional[Union[bool, str, None]] = None,
+        use_auth_token: Optional[Union[bool, str, None]] = None,
         revision: Optional[Union[str, None]] = None,
         force_download: bool = False,
         cache_dir: Optional[str] = None,
@@ -150,7 +148,7 @@ class OnnxRuntimeModel:
         Arguments:
             model_id (`str` or `Path`):
                 Directory from which to load
-            token (`str` or `bool`):
+            use_auth_token (`str` or `bool`):
                 Is needed to load models from a private or gated repository
             revision (`str`):
                 Revision is the specific model version to use. It can be a branch name, a tag name, or a commit id
@@ -181,7 +179,7 @@ class OnnxRuntimeModel:
             model_cache_path = hf_hub_download(
                 repo_id=model_id,
                 filename=model_file_name,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 cache_dir=cache_dir,
                 force_download=force_download,
@@ -192,12 +190,11 @@ class OnnxRuntimeModel:
         return cls(model=model, **kwargs)
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(
         cls,
         model_id: Union[str, Path],
         force_download: bool = True,
-        token: Optional[str] = None,
+        use_auth_token: Optional[str] = None,
         cache_dir: Optional[str] = None,
         **model_kwargs,
     ):
@@ -210,6 +207,6 @@ class OnnxRuntimeModel:
             revision=revision,
             cache_dir=cache_dir,
             force_download=force_download,
-            token=token,
+            use_auth_token=use_auth_token,
             **model_kwargs,
         )

src/diffusers/pipelines/pipeline_flax_utils.py

@@ -24,7 +24,6 @@ import numpy as np
 import PIL.Image
 from flax.core.frozen_dict import FrozenDict
 from huggingface_hub import create_repo, snapshot_download
-from huggingface_hub.utils import validate_hf_hub_args
 from PIL import Image
 from tqdm.auto import tqdm
 
@@ -33,6 +32,7 @@ from ..models.modeling_flax_utils import FLAX_WEIGHTS_NAME, FlaxModelMixin
 from ..schedulers.scheduling_utils_flax import SCHEDULER_CONFIG_NAME, FlaxSchedulerMixin
 from ..utils import (
     CONFIG_NAME,
+    DIFFUSERS_CACHE,
     BaseOutput,
     PushToHubMixin,
     http_user_agent,
@@ -227,7 +227,6 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
                 )
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         r"""
         Instantiate a Flax-based diffusion pipeline from pretrained pipeline weights.
@@ -265,7 +264,7 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -315,11 +314,11 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
         >>> dpm_params["scheduler"] = dpmpp_state
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
-        token = kwargs.pop("token", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         from_pt = kwargs.pop("from_pt", False)
         use_memory_efficient_attention = kwargs.pop("use_memory_efficient_attention", False)
@@ -335,7 +334,7 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
             )
             # make sure we only download sub-folders and `diffusers` filenames
@@ -366,7 +365,7 @@ class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 allow_patterns=allow_patterns,
                 ignore_patterns=ignore_patterns,

src/diffusers/pipelines/pipeline_utils.py

@@ -28,14 +28,7 @@ from typing import Any, Callable, Dict, List, Optional, Union
 import numpy as np
 import PIL.Image
 import torch
-from huggingface_hub import (
-    ModelCard,
-    create_repo,
-    hf_hub_download,
-    model_info,
-    snapshot_download,
-)
-from huggingface_hub.utils import validate_hf_hub_args
+from huggingface_hub import ModelCard, create_repo, hf_hub_download, model_info, snapshot_download
 from packaging import version
 from requests.exceptions import HTTPError
 from tqdm.auto import tqdm
@@ -47,6 +40,8 @@ from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
 from ..utils import (
     CONFIG_NAME,
     DEPRECATED_REVISION_ARGS,
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
     SAFETENSORS_WEIGHTS_NAME,
     WEIGHTS_NAME,
     BaseOutput,
@@ -254,11 +249,10 @@ def variant_compatible_siblings(filenames, variant=None) -> Union[List[os.PathLi
     return usable_filenames, variant_filenames
 
 
-@validate_hf_hub_args
-def warn_deprecated_model_variant(pretrained_model_name_or_path, token, variant, revision, model_filenames):
+def warn_deprecated_model_variant(pretrained_model_name_or_path, use_auth_token, variant, revision, model_filenames):
     info = model_info(
         pretrained_model_name_or_path,
-        token=token,
+        use_auth_token=use_auth_token,
         revision=None,
     )
     filenames = {sibling.rfilename for sibling in info.siblings}
@@ -381,6 +375,7 @@ def _get_pipeline_class(
             custom_pipeline,
             module_file=file_name,
             class_name=class_name,
+            repo_id=repo_id,
             cache_dir=cache_dir,
             revision=revision,
         )
@@ -914,7 +909,6 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         return torch.float32
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         r"""
         Instantiate a PyTorch diffusion pipeline from pretrained pipeline weights.
@@ -982,7 +976,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -1062,12 +1056,12 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         >>> pipeline.scheduler = scheduler
         ```
         """
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         from_flax = kwargs.pop("from_flax", False)
         torch_dtype = kwargs.pop("torch_dtype", None)
@@ -1100,7 +1094,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 force_download=force_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 from_flax=from_flax,
                 use_safetensors=use_safetensors,
@@ -1305,7 +1299,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 "force_download": force_download,
                 "proxies": proxies,
                 "local_files_only": local_files_only,
-                "token": token,
+                "use_auth_token": use_auth_token,
                 "revision": revision,
                 "torch_dtype": torch_dtype,
                 "custom_pipeline": custom_pipeline,
@@ -1535,7 +1529,6 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 cpu_offload(model, device, offload_buffers=offload_buffers)
 
     @classmethod
-    @validate_hf_hub_args
     def download(cls, pretrained_model_name, **kwargs) -> Union[str, os.PathLike]:
         r"""
         Download and cache a PyTorch diffusion pipeline from pretrained pipeline weights.
@@ -1583,7 +1576,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             local_files_only (`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):
@@ -1626,12 +1619,12 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         </Tip>
 
         """
-        cache_dir = kwargs.pop("cache_dir", None)
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
         revision = kwargs.pop("revision", None)
         from_flax = kwargs.pop("from_flax", False)
         custom_pipeline = kwargs.pop("custom_pipeline", None)
@@ -1653,7 +1646,11 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         model_info_call_error: Optional[Exception] = None
         if not local_files_only:
             try:
-                info = model_info(pretrained_model_name, token=token, revision=revision)
+                info = model_info(
+                    pretrained_model_name,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                )
             except HTTPError as e:
                 logger.warn(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.")
                 local_files_only = True
@@ -1668,7 +1665,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 proxies=proxies,
                 force_download=force_download,
                 resume_download=resume_download,
-                token=token,
+                use_auth_token=use_auth_token,
             )
 
             config_dict = cls._dict_from_json_file(config_file)
@@ -1718,7 +1715,9 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             if revision in DEPRECATED_REVISION_ARGS and version.parse(
                 version.parse(__version__).base_version
             ) >= version.parse("0.22.0"):
-                warn_deprecated_model_variant(pretrained_model_name, token, variant, revision, model_filenames)
+                warn_deprecated_model_variant(
+                    pretrained_model_name, use_auth_token, variant, revision, model_filenames
+                )
 
             model_folder_names = {os.path.split(f)[0] for f in model_filenames if os.path.split(f)[0] in folder_names}
 
@@ -1860,7 +1859,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 revision=revision,
                 allow_patterns=allow_patterns,
                 ignore_patterns=ignore_patterns,
@@ -1884,7 +1883,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                         "force_download": force_download,
                         "proxies": proxies,
                         "local_files_only": local_files_only,
-                        "token": token,
+                        "use_auth_token": use_auth_token,
                         "variant": variant,
                         "use_safetensors": use_safetensors,
                     }

src/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py

@@ -447,8 +447,7 @@ def convert_ldm_unet_checkpoint(
 
     # Relevant to StableDiffusionUpscalePipeline
     if "num_class_embeds" in config:
-        if (config["num_class_embeds"] is not None) and ("label_emb.weight" in unet_state_dict):
-            new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"]
+        new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"]
 
     new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
     new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

@@ -34,7 +34,6 @@ from ...loaders import (
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.attention_processor import (
     AttnProcessor2_0,
-    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
@@ -682,6 +681,7 @@ class StableDiffusionXLPipeline(
         add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
         return add_time_ids
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
     def upcast_vae(self):
         dtype = self.vae.dtype
         self.vae.to(dtype=torch.float32)
@@ -692,7 +692,6 @@ class StableDiffusionXLPipeline(
                 XFormersAttnProcessor,
                 LoRAXFormersAttnProcessor,
                 LoRAAttnProcessor2_0,
-                FusedAttnProcessor2_0,
             ),
         )
         # if xformers or torch_2_0 is used attention block does not need
@@ -730,65 +729,6 @@ class StableDiffusionXLPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
-    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
-        """
-        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
-        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        Args:
-            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
-            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
-        """
-        self.fusing_unet = False
-        self.fusing_vae = False
-
-        if unet:
-            self.fusing_unet = True
-            self.unet.fuse_qkv_projections()
-            self.unet.set_attn_processor(FusedAttnProcessor2_0())
-
-        if vae:
-            if not isinstance(self.vae, AutoencoderKL):
-                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
-
-            self.fusing_vae = True
-            self.vae.fuse_qkv_projections()
-            self.vae.set_attn_processor(FusedAttnProcessor2_0())
-
-    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
-        """Disable QKV projection fusion if enabled.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        Args:
-            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
-            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
-
-        """
-        if unet:
-            if not self.fusing_unet:
-                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
-            else:
-                self.unet.unfuse_qkv_projections()
-                self.fusing_unet = False
-
-        if vae:
-            if not self.fusing_vae:
-                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
-            else:
-                self.vae.unfuse_qkv_projections()
-                self.fusing_vae = False
-
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py

@@ -24,7 +24,6 @@ from ...loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, Te
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...models.attention_processor import (
     AttnProcessor2_0,
-    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
@@ -611,7 +610,6 @@ class StableDiffusionXLInstructPix2PixPipeline(
                 XFormersAttnProcessor,
                 LoRAXFormersAttnProcessor,
                 LoRAAttnProcessor2_0,
-                FusedAttnProcessor2_0,
             ),
         )
         # if xformers or torch_2_0 is used attention block does not need

src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py

@@ -10,10 +10,10 @@ from diffusers.utils import deprecate
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...models import ModelMixin
 from ...models.activations import get_activation
+from ...models.attention import Attention
 from ...models.attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
-    Attention,
     AttentionProcessor,
     AttnAddedKVProcessor,
     AttnAddedKVProcessor2_0,
@@ -1000,42 +1000,6 @@ class UNetFlatConditionModel(ModelMixin, ConfigMixin):
                 if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
                     setattr(upsample_block, k, None)
 
-    def fuse_qkv_projections(self):
-        """
-        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
-        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-        """
-        self.original_attn_processors = None
-
-        for _, attn_processor in self.attn_processors.items():
-            if "Added" in str(attn_processor.__class__.__name__):
-                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
-
-        self.original_attn_processors = self.attn_processors
-
-        for module in self.modules():
-            if isinstance(module, Attention):
-                module.fuse_projections(fuse=True)
-
-    def unfuse_qkv_projections(self):
-        """Disables the fused QKV projection if enabled.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        """
-        if self.original_attn_processors is not None:
-            self.set_attn_processor(self.original_attn_processors)
-
     def forward(
         self,
         sample: torch.FloatTensor,

src/diffusers/schedulers/scheduling_deis_multistep.py

@@ -734,16 +734,7 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
             schedule_timesteps = self.timesteps.to(original_samples.device)
             timesteps = timesteps.to(original_samples.device)
 
-        step_indices = []
-        for timestep in timesteps:
-            index_candidates = (schedule_timesteps == timestep).nonzero()
-            if len(index_candidates) == 0:
-                step_index = len(schedule_timesteps) - 1
-            elif len(index_candidates) > 1:
-                step_index = index_candidates[1].item()
-            else:
-                step_index = index_candidates[0].item()
-            step_indices.append(step_index)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
 
         sigma = sigmas[step_indices].flatten()
         while len(sigma.shape) < len(original_samples.shape):

src/diffusers/schedulers/scheduling_dpmsolver_multistep.py

@@ -896,16 +896,7 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             schedule_timesteps = self.timesteps.to(original_samples.device)
             timesteps = timesteps.to(original_samples.device)
 
-        step_indices = []
-        for timestep in timesteps:
-            index_candidates = (schedule_timesteps == timestep).nonzero()
-            if len(index_candidates) == 0:
-                step_index = len(schedule_timesteps) - 1
-            elif len(index_candidates) > 1:
-                step_index = index_candidates[1].item()
-            else:
-                step_index = index_candidates[0].item()
-            step_indices.append(step_index)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
 
         sigma = sigmas[step_indices].flatten()
         while len(sigma.shape) < len(original_samples.shape):

src/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py

@@ -891,16 +891,7 @@ class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin):
             schedule_timesteps = self.timesteps.to(original_samples.device)
             timesteps = timesteps.to(original_samples.device)
 
-        step_indices = []
-        for timestep in timesteps:
-            index_candidates = (schedule_timesteps == timestep).nonzero()
-            if len(index_candidates) == 0:
-                step_index = len(schedule_timesteps) - 1
-            elif len(index_candidates) > 1:
-                step_index = index_candidates[1].item()
-            else:
-                step_index = index_candidates[0].item()
-            step_indices.append(step_index)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
 
         sigma = sigmas[step_indices].flatten()
         while len(sigma.shape) < len(original_samples.shape):

src/diffusers/schedulers/scheduling_dpmsolver_singlestep.py

@@ -897,16 +897,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
             schedule_timesteps = self.timesteps.to(original_samples.device)
             timesteps = timesteps.to(original_samples.device)
 
-        step_indices = []
-        for timestep in timesteps:
-            index_candidates = (schedule_timesteps == timestep).nonzero()
-            if len(index_candidates) == 0:
-                step_index = len(schedule_timesteps) - 1
-            elif len(index_candidates) > 1:
-                step_index = index_candidates[1].item()
-            else:
-                step_index = index_candidates[0].item()
-            step_indices.append(step_index)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
 
         sigma = sigmas[step_indices].flatten()
         while len(sigma.shape) < len(original_samples.shape):

src/diffusers/schedulers/scheduling_euler_discrete.py

@@ -92,43 +92,6 @@ def betas_for_alpha_bar(
     return torch.tensor(betas, dtype=torch.float32)
 
 
-# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
-def rescale_zero_terminal_snr(betas):
-    """
-    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
-
-
-    Args:
-        betas (`torch.FloatTensor`):
-            the betas that the scheduler is being initialized with.
-
-    Returns:
-        `torch.FloatTensor`: rescaled betas with zero terminal SNR
-    """
-    # Convert betas to alphas_bar_sqrt
-    alphas = 1.0 - betas
-    alphas_cumprod = torch.cumprod(alphas, dim=0)
-    alphas_bar_sqrt = alphas_cumprod.sqrt()
-
-    # Store old values.
-    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
-    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
-
-    # Shift so the last timestep is zero.
-    alphas_bar_sqrt -= alphas_bar_sqrt_T
-
-    # Scale so the first timestep is back to the old value.
-    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
-
-    # Convert alphas_bar_sqrt to betas
-    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
-    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
-    alphas = torch.cat([alphas_bar[0:1], alphas])
-    betas = 1 - alphas
-
-    return betas
-
-
 class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
     """
     Euler scheduler.
@@ -165,10 +128,6 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
             An offset added to the inference steps. You can use a combination of `offset=1` and
             `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
             Diffusion.
-        rescale_betas_zero_snr (`bool`, defaults to `False`):
-            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
-            dark samples instead of limiting it to samples with medium brightness. Loosely related to
-            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
     """
 
     _compatibles = [e.name for e in KarrasDiffusionSchedulers]
@@ -190,7 +149,6 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         timestep_spacing: str = "linspace",
         timestep_type: str = "discrete",  # can be "discrete" or "continuous"
         steps_offset: int = 0,
-        rescale_betas_zero_snr: bool = False,
     ):
         if trained_betas is not None:
             self.betas = torch.tensor(trained_betas, dtype=torch.float32)
@@ -205,17 +163,9 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         else:
             raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
 
-        if rescale_betas_zero_snr:
-            self.betas = rescale_zero_terminal_snr(self.betas)
-
         self.alphas = 1.0 - self.betas
         self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
 
-        if rescale_betas_zero_snr:
-            # Close to 0 without being 0 so first sigma is not inf
-            # FP16 smallest positive subnormal works well here
-            self.alphas_cumprod[-1] = 2**-24
-
         sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
         timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy()
 
@@ -241,11 +191,10 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
     @property
     def init_noise_sigma(self):
         # standard deviation of the initial noise distribution
-        max_sigma = max(self.sigmas) if isinstance(self.sigmas, list) else self.sigmas.max()
         if self.config.timestep_spacing in ["linspace", "trailing"]:
-            return max_sigma
+            return self.sigmas.max()
 
-        return (max_sigma**2 + 1) ** 0.5
+        return (self.sigmas.max() ** 2 + 1) ** 0.5
 
     @property
     def step_index(self):
@@ -320,7 +269,7 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         if self.config.interpolation_type == "linear":
             sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
         elif self.config.interpolation_type == "log_linear":
-            sigmas = torch.linspace(np.log(sigmas[-1]), np.log(sigmas[0]), num_inference_steps + 1).exp().numpy()
+            sigmas = torch.linspace(np.log(sigmas[-1]), np.log(sigmas[0]), num_inference_steps + 1).exp()
         else:
             raise ValueError(
                 f"{self.config.interpolation_type} is not implemented. Please specify interpolation_type to either"
@@ -470,9 +419,6 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         if self.step_index is None:
             self._init_step_index(timestep)
 
-        # Upcast to avoid precision issues when computing prev_sample
-        sample = sample.to(torch.float32)
-
         sigma = self.sigmas[self.step_index]
 
         gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
@@ -509,9 +455,6 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
 
         prev_sample = sample + derivative * dt
 
-        # Cast sample back to model compatible dtype
-        prev_sample = prev_sample.to(model_output.dtype)
-
         # upon completion increase step index by one
         self._step_index += 1
 

src/diffusers/schedulers/scheduling_unipc_multistep.py

@@ -828,16 +828,7 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
             schedule_timesteps = self.timesteps.to(original_samples.device)
             timesteps = timesteps.to(original_samples.device)
 
-        step_indices = []
-        for timestep in timesteps:
-            index_candidates = (schedule_timesteps == timestep).nonzero()
-            if len(index_candidates) == 0:
-                step_index = len(schedule_timesteps) - 1
-            elif len(index_candidates) > 1:
-                step_index = index_candidates[1].item()
-            else:
-                step_index = index_candidates[0].item()
-            step_indices.append(step_index)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
 
         sigma = sigmas[step_indices].flatten()
         while len(sigma.shape) < len(original_samples.shape):

src/diffusers/schedulers/scheduling_utils.py

@@ -18,7 +18,6 @@ from enum import Enum
 from typing import Optional, Union
 
 import torch
-from huggingface_hub.utils import validate_hf_hub_args
 
 from ..utils import BaseOutput, PushToHubMixin
 
@@ -82,7 +81,6 @@ class SchedulerMixin(PushToHubMixin):
     has_compatibles = True
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(
         cls,
         pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None,
@@ -122,7 +120,7 @@ class SchedulerMixin(PushToHubMixin):
             local_files_only(`bool`, *optional*, defaults to `False`):
                 Whether to only load local model weights and configuration files or not. If set to `True`, the model
                 won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                 `diffusers-cli login` (stored in `~/.huggingface`) is used.
             revision (`str`, *optional*, defaults to `"main"`):

src/diffusers/schedulers/scheduling_utils_flax.py

@@ -20,7 +20,6 @@ from typing import Optional, Tuple, Union
 
 import flax
 import jax.numpy as jnp
-from huggingface_hub.utils import validate_hf_hub_args
 
 from ..utils import BaseOutput, PushToHubMixin
 
@@ -71,7 +70,6 @@ class FlaxSchedulerMixin(PushToHubMixin):
     has_compatibles = True
 
     @classmethod
-    @validate_hf_hub_args
     def from_pretrained(
         cls,
         pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None,
@@ -112,7 +110,7 @@ class FlaxSchedulerMixin(PushToHubMixin):
                 Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
             local_files_only(`bool`, *optional*, defaults to `False`):
                 Whether or not to only look at local files (i.e., do not try to download the model).
-            token (`str` or *bool*, *optional*):
+            use_auth_token (`str` or *bool*, *optional*):
                 The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
                 when running `transformers-cli login` (stored in `~/.huggingface`).
             revision (`str`, *optional*, defaults to `"main"`):

src/diffusers/utils/__init__.py

@@ -21,6 +21,7 @@ from .. import __version__
 from .constants import (
     CONFIG_NAME,
     DEPRECATED_REVISION_ARGS,
+    DIFFUSERS_CACHE,
     DIFFUSERS_DYNAMIC_MODULE_NAME,
     FLAX_WEIGHTS_NAME,
     HF_MODULES_CACHE,
@@ -37,6 +38,7 @@ from .doc_utils import replace_example_docstring
 from .dynamic_modules_utils import get_class_from_dynamic_module
 from .export_utils import export_to_gif, export_to_obj, export_to_ply, export_to_video
 from .hub_utils import (
+    HF_HUB_OFFLINE,
     PushToHubMixin,
     _add_variant,
     _get_model_file,

src/diffusers/utils/constants.py

@@ -14,13 +14,15 @@
 import importlib
 import os
 
-from huggingface_hub.constants import HF_HOME
+from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE, hf_cache_home
 from packaging import version
 
 from ..dependency_versions_check import dep_version_check
 from .import_utils import ENV_VARS_TRUE_VALUES, is_peft_available, is_transformers_available
 
 
+default_cache_path = HUGGINGFACE_HUB_CACHE
+
 MIN_PEFT_VERSION = "0.6.0"
 MIN_TRANSFORMERS_VERSION = "4.34.0"
 _CHECK_PEFT = os.environ.get("_CHECK_PEFT", "1") in ENV_VARS_TRUE_VALUES
@@ -33,8 +35,9 @@ ONNX_WEIGHTS_NAME = "model.onnx"
 SAFETENSORS_WEIGHTS_NAME = "diffusion_pytorch_model.safetensors"
 ONNX_EXTERNAL_WEIGHTS_NAME = "weights.pb"
 HUGGINGFACE_CO_RESOLVE_ENDPOINT = os.environ.get("HF_ENDPOINT", "https://huggingface.co")
+DIFFUSERS_CACHE = default_cache_path
 DIFFUSERS_DYNAMIC_MODULE_NAME = "diffusers_modules"
-HF_MODULES_CACHE = os.getenv("HF_MODULES_CACHE", os.path.join(HF_HOME, "modules"))
+HF_MODULES_CACHE = os.getenv("HF_MODULES_CACHE", os.path.join(hf_cache_home, "modules"))
 DEPRECATED_REVISION_ARGS = ["fp16", "non-ema"]
 
 # Below should be `True` if the current version of `peft` and `transformers` are compatible with

src/diffusers/utils/dummy_pt_objects.py

@@ -92,21 +92,6 @@ class ControlNetModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
-class ControlNetXSModel(metaclass=DummyObject):
-    _backends = ["torch"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch"])
-
-
 class Kandinsky3UNet(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -737,21 +737,6 @@ class StableDiffusionControlNetPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
-class StableDiffusionControlNetXSPipeline(metaclass=DummyObject):
-    _backends = ["torch", "transformers"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch", "transformers"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-
 class StableDiffusionDepth2ImgPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 
@@ -1097,21 +1082,6 @@ class StableDiffusionXLControlNetPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
-class StableDiffusionXLControlNetXSPipeline(metaclass=DummyObject):
-    _backends = ["torch", "transformers"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch", "transformers"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-
 class StableDiffusionXLImg2ImgPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

src/diffusers/utils/dynamic_modules_utils.py

@@ -25,8 +25,7 @@ from pathlib import Path
 from typing import Dict, Optional, Union
 from urllib import request
 
-from huggingface_hub import cached_download, hf_hub_download, model_info
-from huggingface_hub.utils import validate_hf_hub_args
+from huggingface_hub import HfFolder, cached_download, hf_hub_download, model_info
 from packaging import version
 
 from .. import __version__
@@ -195,7 +194,6 @@ def find_pipeline_class(loaded_module):
     return pipeline_class
 
 
-@validate_hf_hub_args
 def get_cached_module_file(
     pretrained_model_name_or_path: Union[str, os.PathLike],
     module_file: str,
@@ -203,7 +201,7 @@ def get_cached_module_file(
     force_download: bool = False,
     resume_download: bool = False,
     proxies: Optional[Dict[str, str]] = None,
-    token: Optional[Union[bool, str]] = None,
+    use_auth_token: Optional[Union[bool, str]] = None,
     revision: Optional[str] = None,
     local_files_only: bool = False,
 ):
@@ -234,7 +232,7 @@ def get_cached_module_file(
         proxies (`Dict[str, str]`, *optional*):
             A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
             'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
-        token (`str` or *bool*, *optional*):
+        use_auth_token (`str` or *bool*, *optional*):
             The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
             when running `transformers-cli login` (stored in `~/.huggingface`).
         revision (`str`, *optional*, defaults to `"main"`):
@@ -246,7 +244,7 @@ def get_cached_module_file(
 
     <Tip>
 
-    You may pass a token in `token` if you are not logged in (`huggingface-cli login`) and want to use private
+    You may pass a token in `use_auth_token` if you are not logged in (`huggingface-cli long`) and want to use private
     or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models).
 
     </Tip>
@@ -291,7 +289,7 @@ def get_cached_module_file(
                 proxies=proxies,
                 resume_download=resume_download,
                 local_files_only=local_files_only,
-                token=False,
+                use_auth_token=False,
             )
             submodule = "git"
             module_file = pretrained_model_name_or_path + ".py"
@@ -309,7 +307,7 @@ def get_cached_module_file(
                 proxies=proxies,
                 resume_download=resume_download,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
             )
             submodule = os.path.join("local", "--".join(pretrained_model_name_or_path.split("/")))
         except EnvironmentError:
@@ -334,6 +332,13 @@ def get_cached_module_file(
     else:
         # Get the commit hash
         # TODO: we will get this info in the etag soon, so retrieve it from there and not here.
+        if isinstance(use_auth_token, str):
+            token = use_auth_token
+        elif use_auth_token is True:
+            token = HfFolder.get_token()
+        else:
+            token = None
+
         commit_hash = model_info(pretrained_model_name_or_path, revision=revision, token=token).sha
 
         # The module file will end up being placed in a subfolder with the git hash of the repo. This way we get the
@@ -354,14 +359,13 @@ def get_cached_module_file(
                     force_download=force_download,
                     resume_download=resume_download,
                     proxies=proxies,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     revision=revision,
                     local_files_only=local_files_only,
                 )
     return os.path.join(full_submodule, module_file)
 
 
-@validate_hf_hub_args
 def get_class_from_dynamic_module(
     pretrained_model_name_or_path: Union[str, os.PathLike],
     module_file: str,
@@ -370,7 +374,7 @@ def get_class_from_dynamic_module(
     force_download: bool = False,
     resume_download: bool = False,
     proxies: Optional[Dict[str, str]] = None,
-    token: Optional[Union[bool, str]] = None,
+    use_auth_token: Optional[Union[bool, str]] = None,
     revision: Optional[str] = None,
     local_files_only: bool = False,
     **kwargs,
@@ -410,7 +414,7 @@ def get_class_from_dynamic_module(
         proxies (`Dict[str, str]`, *optional*):
             A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
             'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
-        token (`str` or `bool`, *optional*):
+        use_auth_token (`str` or `bool`, *optional*):
             The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
             when running `transformers-cli login` (stored in `~/.huggingface`).
         revision (`str`, *optional*, defaults to `"main"`):
@@ -422,7 +426,7 @@ def get_class_from_dynamic_module(
 
     <Tip>
 
-    You may pass a token in `token` if you are not logged in (`huggingface-cli login`) and want to use private
+    You may pass a token in `use_auth_token` if you are not logged in (`huggingface-cli long`) and want to use private
     or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models).
 
     </Tip>
@@ -445,7 +449,7 @@ def get_class_from_dynamic_module(
         force_download=force_download,
         resume_download=resume_download,
         proxies=proxies,
-        token=token,
+        use_auth_token=use_auth_token,
         revision=revision,
         local_files_only=local_files_only,
     )

src/diffusers/utils/hub_utils.py

@@ -25,21 +25,20 @@ from typing import Dict, Optional, Union
 from uuid import uuid4
 
 from huggingface_hub import (
+    HfFolder,
     ModelCard,
     ModelCardData,
     create_repo,
-    get_full_repo_name,
     hf_hub_download,
     upload_folder,
+    whoami,
 )
-from huggingface_hub.constants import HF_HUB_CACHE, HF_HUB_DISABLE_TELEMETRY, HF_HUB_OFFLINE
 from huggingface_hub.file_download import REGEX_COMMIT_HASH
 from huggingface_hub.utils import (
     EntryNotFoundError,
     RepositoryNotFoundError,
     RevisionNotFoundError,
     is_jinja_available,
-    validate_hf_hub_args,
 )
 from packaging import version
 from requests import HTTPError
@@ -47,6 +46,7 @@ from requests import HTTPError
 from .. import __version__
 from .constants import (
     DEPRECATED_REVISION_ARGS,
+    DIFFUSERS_CACHE,
     HUGGINGFACE_CO_RESOLVE_ENDPOINT,
     SAFETENSORS_WEIGHTS_NAME,
     WEIGHTS_NAME,
@@ -69,6 +69,9 @@ logger = get_logger(__name__)
 
 MODEL_CARD_TEMPLATE_PATH = Path(__file__).parent / "model_card_template.md"
 SESSION_ID = uuid4().hex
+HF_HUB_OFFLINE = os.getenv("HF_HUB_OFFLINE", "").upper() in ENV_VARS_TRUE_VALUES
+DISABLE_TELEMETRY = os.getenv("DISABLE_TELEMETRY", "").upper() in ENV_VARS_TRUE_VALUES
+HUGGINGFACE_CO_TELEMETRY = HUGGINGFACE_CO_RESOLVE_ENDPOINT + "/api/telemetry/"
 
 
 def http_user_agent(user_agent: Union[Dict, str, None] = None) -> str:
@@ -76,7 +79,7 @@ def http_user_agent(user_agent: Union[Dict, str, None] = None) -> str:
     Formats a user-agent string with basic info about a request.
     """
     ua = f"diffusers/{__version__}; python/{sys.version.split()[0]}; session_id/{SESSION_ID}"
-    if HF_HUB_DISABLE_TELEMETRY or HF_HUB_OFFLINE:
+    if DISABLE_TELEMETRY or HF_HUB_OFFLINE:
         return ua + "; telemetry/off"
     if is_torch_available():
         ua += f"; torch/{_torch_version}"
@@ -95,6 +98,16 @@ def http_user_agent(user_agent: Union[Dict, str, None] = None) -> str:
     return ua
 
 
+def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
+    if token is None:
+        token = HfFolder.get_token()
+    if organization is None:
+        username = whoami(token)["name"]
+        return f"{username}/{model_id}"
+    else:
+        return f"{organization}/{model_id}"
+
+
 def create_model_card(args, model_name):
     if not is_jinja_available():
         raise ValueError(
@@ -170,7 +183,7 @@ old_diffusers_cache = os.path.join(hf_cache_home, "diffusers")
 
 def move_cache(old_cache_dir: Optional[str] = None, new_cache_dir: Optional[str] = None) -> None:
     if new_cache_dir is None:
-        new_cache_dir = HF_HUB_CACHE
+        new_cache_dir = DIFFUSERS_CACHE
     if old_cache_dir is None:
         old_cache_dir = old_diffusers_cache
 
@@ -190,7 +203,7 @@ def move_cache(old_cache_dir: Optional[str] = None, new_cache_dir: Optional[str]
     # At this point, old_cache_dir contains symlinks to the new cache (it can still be used).
 
 
-cache_version_file = os.path.join(HF_HUB_CACHE, "version_diffusers_cache.txt")
+cache_version_file = os.path.join(DIFFUSERS_CACHE, "version_diffusers_cache.txt")
 if not os.path.isfile(cache_version_file):
     cache_version = 0
 else:
@@ -220,12 +233,12 @@ if cache_version < 1:
 
 if cache_version < 1:
     try:
-        os.makedirs(HF_HUB_CACHE, exist_ok=True)
+        os.makedirs(DIFFUSERS_CACHE, exist_ok=True)
         with open(cache_version_file, "w") as f:
             f.write("1")
     except Exception:
         logger.warning(
-            f"There was a problem when trying to write in your cache folder ({HF_HUB_CACHE}). Please, ensure "
+            f"There was a problem when trying to write in your cache folder ({DIFFUSERS_CACHE}). Please, ensure "
             "the directory exists and can be written to."
         )
 
@@ -239,21 +252,20 @@ def _add_variant(weights_name: str, variant: Optional[str] = None) -> str:
     return weights_name
 
 
-@validate_hf_hub_args
 def _get_model_file(
-    pretrained_model_name_or_path: Union[str, Path],
+    pretrained_model_name_or_path,
     *,
-    weights_name: str,
-    subfolder: Optional[str],
-    cache_dir: Optional[str],
-    force_download: bool,
-    proxies: Optional[Dict],
-    resume_download: bool,
-    local_files_only: bool,
-    token: Optional[str],
-    user_agent: Union[Dict, str, None],
-    revision: Optional[str],
-    commit_hash: Optional[str] = None,
+    weights_name,
+    subfolder,
+    cache_dir,
+    force_download,
+    proxies,
+    resume_download,
+    local_files_only,
+    use_auth_token,
+    user_agent,
+    revision,
+    commit_hash=None,
 ):
     pretrained_model_name_or_path = str(pretrained_model_name_or_path)
     if os.path.isfile(pretrained_model_name_or_path):
@@ -288,7 +300,7 @@ def _get_model_file(
                     proxies=proxies,
                     resume_download=resume_download,
                     local_files_only=local_files_only,
-                    token=token,
+                    use_auth_token=use_auth_token,
                     user_agent=user_agent,
                     subfolder=subfolder,
                     revision=revision or commit_hash,
@@ -313,7 +325,7 @@ def _get_model_file(
                 proxies=proxies,
                 resume_download=resume_download,
                 local_files_only=local_files_only,
-                token=token,
+                use_auth_token=use_auth_token,
                 user_agent=user_agent,
                 subfolder=subfolder,
                 revision=revision or commit_hash,
@@ -324,7 +336,7 @@ def _get_model_file(
             raise EnvironmentError(
                 f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier "
                 "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a "
-                "token having permission to this repo with `token` or log in with `huggingface-cli "
+                "token having permission to this repo with `use_auth_token` or log in with `huggingface-cli "
                 "login`."
             )
         except RevisionNotFoundError:

src/diffusers/utils/testing_utils.py

@@ -17,7 +17,7 @@ from contextlib import contextmanager
 from distutils.util import strtobool
 from io import BytesIO, StringIO
 from pathlib import Path
-from typing import Callable, Dict, List, Optional, Union
+from typing import List, Optional, Union
 
 import numpy as np
 import PIL.Image
@@ -58,17 +58,6 @@ USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version
 if is_torch_available():
     import torch
 
-    # Set a backend environment variable for any extra module import required for a custom accelerator
-    if "DIFFUSERS_TEST_BACKEND" in os.environ:
-        backend = os.environ["DIFFUSERS_TEST_BACKEND"]
-        try:
-            _ = importlib.import_module(backend)
-        except ModuleNotFoundError as e:
-            raise ModuleNotFoundError(
-                f"Failed to import `DIFFUSERS_TEST_BACKEND` '{backend}'! This should be the name of an installed module \
-                    to enable a specified backend.):\n{e}"
-            ) from e
-
     if "DIFFUSERS_TEST_DEVICE" in os.environ:
         torch_device = os.environ["DIFFUSERS_TEST_DEVICE"]
         try:
@@ -221,36 +210,6 @@ def require_torch_gpu(test_case):
     )
 
 
-# These decorators are for accelerator-specific behaviours that are not GPU-specific
-def require_torch_accelerator(test_case):
-    """Decorator marking a test that requires an accelerator backend and PyTorch."""
-    return unittest.skipUnless(is_torch_available() and torch_device != "cpu", "test requires accelerator+PyTorch")(
-        test_case
-    )
-
-
-def require_torch_accelerator_with_fp16(test_case):
-    """Decorator marking a test that requires an accelerator with support for the FP16 data type."""
-    return unittest.skipUnless(_is_torch_fp16_available(torch_device), "test requires accelerator with fp16 support")(
-        test_case
-    )
-
-
-def require_torch_accelerator_with_fp64(test_case):
-    """Decorator marking a test that requires an accelerator with support for the FP64 data type."""
-    return unittest.skipUnless(_is_torch_fp64_available(torch_device), "test requires accelerator with fp64 support")(
-        test_case
-    )
-
-
-def require_torch_accelerator_with_training(test_case):
-    """Decorator marking a test that requires an accelerator with support for training."""
-    return unittest.skipUnless(
-        is_torch_available() and backend_supports_training(torch_device),
-        "test requires accelerator with training support",
-    )(test_case)
-
-
 def skip_mps(test_case):
     """Decorator marking a test to skip if torch_device is 'mps'"""
     return unittest.skipUnless(torch_device != "mps", "test requires non 'mps' device")(test_case)
@@ -807,139 +766,3 @@ def disable_full_determinism():
     os.environ["CUDA_LAUNCH_BLOCKING"] = "0"
     os.environ["CUBLAS_WORKSPACE_CONFIG"] = ""
     torch.use_deterministic_algorithms(False)
-
-
-# Utils for custom and alternative accelerator devices
-def _is_torch_fp16_available(device):
-    if not is_torch_available():
-        return False
-
-    import torch
-
-    device = torch.device(device)
-
-    try:
-        x = torch.zeros((2, 2), dtype=torch.float16).to(device)
-        _ = x @ x
-    except Exception as e:
-        if device.type == "cuda":
-            raise ValueError(
-                f"You have passed a device of type 'cuda' which should work with 'fp16', but 'cuda' does not seem to be correctly installed on your machine: {e}"
-            )
-
-        return False
-
-
-def _is_torch_fp64_available(device):
-    if not is_torch_available():
-        return False
-
-    import torch
-
-    try:
-        x = torch.zeros((2, 2), dtype=torch.float64).to(device)
-        _ = x @ x
-    except Exception as e:
-        if device.type == "cuda":
-            raise ValueError(
-                f"You have passed a device of type 'cuda' which should work with 'fp64', but 'cuda' does not seem to be correctly installed on your machine: {e}"
-            )
-
-        return False
-
-
-# Guard these lookups for when Torch is not used - alternative accelerator support is for PyTorch
-if is_torch_available():
-    # Behaviour flags
-    BACKEND_SUPPORTS_TRAINING = {"cuda": True, "cpu": True, "mps": False, "default": True}
-
-    # Function definitions
-    BACKEND_EMPTY_CACHE = {"cuda": torch.cuda.empty_cache, "cpu": None, "mps": None, "default": None}
-    BACKEND_DEVICE_COUNT = {"cuda": torch.cuda.device_count, "cpu": lambda: 0, "mps": lambda: 0, "default": 0}
-    BACKEND_MANUAL_SEED = {"cuda": torch.cuda.manual_seed, "cpu": torch.manual_seed, "default": torch.manual_seed}
-
-
-# This dispatches a defined function according to the accelerator from the function definitions.
-def _device_agnostic_dispatch(device: str, dispatch_table: Dict[str, Callable], *args, **kwargs):
-    if device not in dispatch_table:
-        return dispatch_table["default"](*args, **kwargs)
-
-    fn = dispatch_table[device]
-
-    # Some device agnostic functions return values. Need to guard against 'None' instead at
-    # user level
-    if fn is None:
-        return None
-
-    return fn(*args, **kwargs)
-
-
-# These are callables which automatically dispatch the function specific to the accelerator
-def backend_manual_seed(device: str, seed: int):
-    return _device_agnostic_dispatch(device, BACKEND_MANUAL_SEED, seed)
-
-
-def backend_empty_cache(device: str):
-    return _device_agnostic_dispatch(device, BACKEND_EMPTY_CACHE)
-
-
-def backend_device_count(device: str):
-    return _device_agnostic_dispatch(device, BACKEND_DEVICE_COUNT)
-
-
-# These are callables which return boolean behaviour flags and can be used to specify some
-# device agnostic alternative where the feature is unsupported.
-def backend_supports_training(device: str):
-    if not is_torch_available():
-        return False
-
-    if device not in BACKEND_SUPPORTS_TRAINING:
-        device = "default"
-
-    return BACKEND_SUPPORTS_TRAINING[device]
-
-
-# Guard for when Torch is not available
-if is_torch_available():
-    # Update device function dict mapping
-    def update_mapping_from_spec(device_fn_dict: Dict[str, Callable], attribute_name: str):
-        try:
-            # Try to import the function directly
-            spec_fn = getattr(device_spec_module, attribute_name)
-            device_fn_dict[torch_device] = spec_fn
-        except AttributeError as e:
-            # If the function doesn't exist, and there is no default, throw an error
-            if "default" not in device_fn_dict:
-                raise AttributeError(
-                    f"`{attribute_name}` not found in '{device_spec_path}' and no default fallback function found."
-                ) from e
-
-    if "DIFFUSERS_TEST_DEVICE_SPEC" in os.environ:
-        device_spec_path = os.environ["DIFFUSERS_TEST_DEVICE_SPEC"]
-        if not Path(device_spec_path).is_file():
-            raise ValueError(f"Specified path to device specification file is not found. Received {device_spec_path}")
-
-        try:
-            import_name = device_spec_path[: device_spec_path.index(".py")]
-        except ValueError as e:
-            raise ValueError(f"Provided device spec file is not a Python file! Received {device_spec_path}") from e
-
-        device_spec_module = importlib.import_module(import_name)
-
-        try:
-            device_name = device_spec_module.DEVICE_NAME
-        except AttributeError:
-            raise AttributeError("Device spec file did not contain `DEVICE_NAME`")
-
-        if "DIFFUSERS_TEST_DEVICE" in os.environ and torch_device != device_name:
-            msg = f"Mismatch between environment variable `DIFFUSERS_TEST_DEVICE` '{torch_device}' and device found in spec '{device_name}'\n"
-            msg += "Either unset `DIFFUSERS_TEST_DEVICE` or ensure it matches device spec name."
-            raise ValueError(msg)
-
-        torch_device = device_name
-
-        # Add one entry here for each `BACKEND_*` dictionary.
-        update_mapping_from_spec(BACKEND_MANUAL_SEED, "MANUAL_SEED_FN")
-        update_mapping_from_spec(BACKEND_EMPTY_CACHE, "EMPTY_CACHE_FN")
-        update_mapping_from_spec(BACKEND_DEVICE_COUNT, "DEVICE_COUNT_FN")
-        update_mapping_from_spec(BACKEND_SUPPORTS_TRAINING, "SUPPORTS_TRAINING")

tests/models/test_layers_utils.py

@@ -25,11 +25,7 @@ from diffusers.models.embeddings import get_timestep_embedding
 from diffusers.models.lora import LoRACompatibleLinear
 from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
 from diffusers.models.transformer_2d import Transformer2DModel
-from diffusers.utils.testing_utils import (
-    backend_manual_seed,
-    require_torch_accelerator_with_fp64,
-    torch_device,
-)
+from diffusers.utils.testing_utils import torch_device
 
 
 class EmbeddingsTests(unittest.TestCase):
@@ -319,7 +315,8 @@ class ResnetBlock2DTests(unittest.TestCase):
 class Transformer2DModelTests(unittest.TestCase):
     def test_spatial_transformer_default(self):
         torch.manual_seed(0)
-        backend_manual_seed(torch_device, 0)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
         sample = torch.randn(1, 32, 64, 64).to(torch_device)
         spatial_transformer_block = Transformer2DModel(
@@ -342,7 +339,8 @@ class Transformer2DModelTests(unittest.TestCase):
 
     def test_spatial_transformer_cross_attention_dim(self):
         torch.manual_seed(0)
-        backend_manual_seed(torch_device, 0)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
         sample = torch.randn(1, 64, 64, 64).to(torch_device)
         spatial_transformer_block = Transformer2DModel(
@@ -365,7 +363,8 @@ class Transformer2DModelTests(unittest.TestCase):
 
     def test_spatial_transformer_timestep(self):
         torch.manual_seed(0)
-        backend_manual_seed(torch_device, 0)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
         num_embeds_ada_norm = 5
 
@@ -402,7 +401,8 @@ class Transformer2DModelTests(unittest.TestCase):
 
     def test_spatial_transformer_dropout(self):
         torch.manual_seed(0)
-        backend_manual_seed(torch_device, 0)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
         sample = torch.randn(1, 32, 64, 64).to(torch_device)
         spatial_transformer_block = (
@@ -427,10 +427,11 @@ class Transformer2DModelTests(unittest.TestCase):
         )
         assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
 
-    @require_torch_accelerator_with_fp64
+    @unittest.skipIf(torch_device == "mps", "MPS does not support float64")
     def test_spatial_transformer_discrete(self):
         torch.manual_seed(0)
-        backend_manual_seed(torch_device, 0)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
         num_embed = 5
 

tests/models/test_modeling_common.py

@@ -35,7 +35,6 @@ from diffusers.utils.testing_utils import (
     CaptureLogger,
     require_python39_or_higher,
     require_torch_2,
-    require_torch_accelerator_with_training,
     require_torch_gpu,
     run_test_in_subprocess,
     torch_device,
@@ -537,7 +536,7 @@ class ModelTesterMixin:
 
         self.assertEqual(output_1.shape, output_2.shape)
 
-    @require_torch_accelerator_with_training
+    @unittest.skipIf(torch_device == "mps", "Training is not supported in mps")
     def test_training(self):
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
 
@@ -554,7 +553,7 @@ class ModelTesterMixin:
         loss = torch.nn.functional.mse_loss(output, noise)
         loss.backward()
 
-    @require_torch_accelerator_with_training
+    @unittest.skipIf(torch_device == "mps", "Training is not supported in mps")
     def test_ema_training(self):
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
 
@@ -625,7 +624,7 @@ class ModelTesterMixin:
 
         recursive_check(outputs_tuple, outputs_dict)
 
-    @require_torch_accelerator_with_training
+    @unittest.skipIf(torch_device == "mps", "Gradient checkpointing skipped on MPS")
     def test_enable_disable_gradient_checkpointing(self):
         if not self.model_class._supports_gradient_checkpointing:
             return  # Skip test if model does not support gradient checkpointing

tests/models/test_models_prior.py

@@ -21,14 +21,7 @@ import torch
 from parameterized import parameterized
 
 from diffusers import PriorTransformer
-from diffusers.utils.testing_utils import (
-    backend_empty_cache,
-    enable_full_determinism,
-    floats_tensor,
-    slow,
-    torch_all_close,
-    torch_device,
-)
+from diffusers.utils.testing_utils import enable_full_determinism, floats_tensor, slow, torch_all_close, torch_device
 
 from .test_modeling_common import ModelTesterMixin
 
@@ -164,7 +157,7 @@ class PriorTransformerIntegrationTests(unittest.TestCase):
         # clean up the VRAM after each test
         super().tearDown()
         gc.collect()
-        backend_empty_cache(torch_device)
+        torch.cuda.empty_cache()
 
     @parameterized.expand(
         [

tests/models/test_models_unet_1d.py

@@ -18,12 +18,7 @@ import unittest
 import torch
 
 from diffusers import UNet1DModel
-from diffusers.utils.testing_utils import (
-    backend_manual_seed,
-    floats_tensor,
-    slow,
-    torch_device,
-)
+from diffusers.utils.testing_utils import floats_tensor, slow, torch_device
 
 from .test_modeling_common import ModelTesterMixin, UNetTesterMixin
 
@@ -108,7 +103,8 @@ class UNet1DModelTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
     def test_output_pretrained(self):
         model = UNet1DModel.from_pretrained("bglick13/hopper-medium-v2-value-function-hor32", subfolder="unet")
         torch.manual_seed(0)
-        backend_manual_seed(torch_device, 0)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
         num_features = model.config.in_channels
         seq_len = 16
@@ -248,7 +244,8 @@ class UNetRLModelTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
             "bglick13/hopper-medium-v2-value-function-hor32", output_loading_info=True, subfolder="value_function"
         )
         torch.manual_seed(0)
-        backend_manual_seed(torch_device, 0)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
         num_features = value_function.config.in_channels
         seq_len = 14

tests/models/test_models_unet_2d.py

@@ -24,7 +24,6 @@ from diffusers.utils import logging
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
     floats_tensor,
-    require_torch_accelerator,
     slow,
     torch_all_close,
     torch_device,
@@ -154,7 +153,7 @@ class UNetLDMModelTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
 
         assert image is not None, "Make sure output is not None"
 
-    @require_torch_accelerator
+    @unittest.skipIf(torch_device != "cuda", "This test is supposed to run on GPU")
     def test_from_pretrained_accelerate(self):
         model, _ = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update", output_loading_info=True)
         model.to(torch_device)
@@ -162,7 +161,7 @@ class UNetLDMModelTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
 
         assert image is not None, "Make sure output is not None"
 
-    @require_torch_accelerator
+    @unittest.skipIf(torch_device != "cuda", "This test is supposed to run on GPU")
     def test_from_pretrained_accelerate_wont_change_results(self):
         # by defautl model loading will use accelerate as `low_cpu_mem_usage=True`
         model_accelerate, _ = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update", output_loading_info=True)

tests/models/test_models_unet_2d_condition.py

@@ -30,15 +30,10 @@ from diffusers.models.embeddings import ImageProjection, Resampler
 from diffusers.utils import logging
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import (
-    backend_empty_cache,
     enable_full_determinism,
     floats_tensor,
     load_hf_numpy,
-    require_torch_accelerator,
-    require_torch_accelerator_with_fp16,
-    require_torch_accelerator_with_training,
     require_torch_gpu,
-    skip_mps,
     slow,
     torch_all_close,
     torch_device,
@@ -285,7 +280,7 @@ class UNet2DConditionModelTests(ModelTesterMixin, UNetTesterMixin, unittest.Test
             == "XFormersAttnProcessor"
         ), "xformers is not enabled"
 
-    @require_torch_accelerator_with_training
+    @unittest.skipIf(torch_device == "mps", "Gradient checkpointing skipped on MPS")
     def test_gradient_checkpointing(self):
         # enable deterministic behavior for gradient checkpointing
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
@@ -869,7 +864,7 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
         # clean up the VRAM after each test
         super().tearDown()
         gc.collect()
-        backend_empty_cache(torch_device)
+        torch.cuda.empty_cache()
 
     def get_latents(self, seed=0, shape=(4, 4, 64, 64), fp16=False):
         dtype = torch.float16 if fp16 else torch.float32
@@ -887,7 +882,6 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
 
         return model
 
-    @require_torch_gpu
     def test_set_attention_slice_auto(self):
         torch.cuda.empty_cache()
         torch.cuda.reset_max_memory_allocated()
@@ -907,7 +901,6 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
 
         assert mem_bytes < 5 * 10**9
 
-    @require_torch_gpu
     def test_set_attention_slice_max(self):
         torch.cuda.empty_cache()
         torch.cuda.reset_max_memory_allocated()
@@ -927,7 +920,6 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
 
         assert mem_bytes < 5 * 10**9
 
-    @require_torch_gpu
     def test_set_attention_slice_int(self):
         torch.cuda.empty_cache()
         torch.cuda.reset_max_memory_allocated()
@@ -947,7 +939,6 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
 
         assert mem_bytes < 5 * 10**9
 
-    @require_torch_gpu
     def test_set_attention_slice_list(self):
         torch.cuda.empty_cache()
         torch.cuda.reset_max_memory_allocated()
@@ -984,7 +975,7 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator_with_fp16
+    @require_torch_gpu
     def test_compvis_sd_v1_4(self, seed, timestep, expected_slice):
         model = self.get_unet_model(model_id="CompVis/stable-diffusion-v1-4")
         latents = self.get_latents(seed)
@@ -1012,7 +1003,7 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator_with_fp16
+    @require_torch_gpu
     def test_compvis_sd_v1_4_fp16(self, seed, timestep, expected_slice):
         model = self.get_unet_model(model_id="CompVis/stable-diffusion-v1-4", fp16=True)
         latents = self.get_latents(seed, fp16=True)
@@ -1040,8 +1031,7 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator
-    @skip_mps
+    @require_torch_gpu
     def test_compvis_sd_v1_5(self, seed, timestep, expected_slice):
         model = self.get_unet_model(model_id="runwayml/stable-diffusion-v1-5")
         latents = self.get_latents(seed)
@@ -1069,7 +1059,7 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator_with_fp16
+    @require_torch_gpu
     def test_compvis_sd_v1_5_fp16(self, seed, timestep, expected_slice):
         model = self.get_unet_model(model_id="runwayml/stable-diffusion-v1-5", fp16=True)
         latents = self.get_latents(seed, fp16=True)
@@ -1097,8 +1087,7 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator
-    @skip_mps
+    @require_torch_gpu
     def test_compvis_sd_inpaint(self, seed, timestep, expected_slice):
         model = self.get_unet_model(model_id="runwayml/stable-diffusion-inpainting")
         latents = self.get_latents(seed, shape=(4, 9, 64, 64))
@@ -1126,7 +1115,7 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator_with_fp16
+    @require_torch_gpu
     def test_compvis_sd_inpaint_fp16(self, seed, timestep, expected_slice):
         model = self.get_unet_model(model_id="runwayml/stable-diffusion-inpainting", fp16=True)
         latents = self.get_latents(seed, shape=(4, 9, 64, 64), fp16=True)
@@ -1154,7 +1143,7 @@ class UNet2DConditionModelIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator_with_fp16
+    @require_torch_gpu
     def test_stabilityai_sd_v2_fp16(self, seed, timestep, expected_slice):
         model = self.get_unet_model(model_id="stabilityai/stable-diffusion-2", fp16=True)
         latents = self.get_latents(seed, shape=(4, 4, 96, 96), fp16=True)

tests/models/test_models_vae.py

@@ -31,15 +31,10 @@ from diffusers import (
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.loading_utils import load_image
 from diffusers.utils.testing_utils import (
-    backend_empty_cache,
     enable_full_determinism,
     floats_tensor,
     load_hf_numpy,
-    require_torch_accelerator,
-    require_torch_accelerator_with_fp16,
-    require_torch_accelerator_with_training,
     require_torch_gpu,
-    skip_mps,
     slow,
     torch_all_close,
     torch_device,
@@ -162,7 +157,7 @@ class AutoencoderKLTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
     def test_training(self):
         pass
 
-    @require_torch_accelerator_with_training
+    @unittest.skipIf(torch_device == "mps", "Gradient checkpointing skipped on MPS")
     def test_gradient_checkpointing(self):
         # enable deterministic behavior for gradient checkpointing
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
@@ -218,12 +213,10 @@ class AutoencoderKLTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
         model = model.to(torch_device)
         model.eval()
 
-        # Keep generator on CPU for non-CUDA devices to compare outputs with CPU result tensors
-        generator_device = "cpu" if not torch_device.startswith("cuda") else "cuda"
-        if torch_device != "mps":
-            generator = torch.Generator(device=generator_device).manual_seed(0)
-        else:
+        if torch_device == "mps":
             generator = torch.manual_seed(0)
+        else:
+            generator = torch.Generator(device=torch_device).manual_seed(0)
 
         image = torch.randn(
             1,
@@ -254,7 +247,7 @@ class AutoencoderKLTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
                     -9.8644e-03,
                 ]
             )
-        elif generator_device == "cpu":
+        elif torch_device == "cpu":
             expected_output_slice = torch.tensor(
                 [
                     -0.1352,
@@ -485,7 +478,7 @@ class AutoencoderTinyIntegrationTests(unittest.TestCase):
         # clean up the VRAM after each test
         super().tearDown()
         gc.collect()
-        backend_empty_cache(torch_device)
+        torch.cuda.empty_cache()
 
     def get_file_format(self, seed, shape):
         return f"gaussian_noise_s={seed}_shape={'_'.join([str(s) for s in shape])}.npy"
@@ -565,7 +558,7 @@ class AutoencoderKLIntegrationTests(unittest.TestCase):
         # clean up the VRAM after each test
         super().tearDown()
         gc.collect()
-        backend_empty_cache(torch_device)
+        torch.cuda.empty_cache()
 
     def get_sd_image(self, seed=0, shape=(4, 3, 512, 512), fp16=False):
         dtype = torch.float16 if fp16 else torch.float32
@@ -587,10 +580,9 @@ class AutoencoderKLIntegrationTests(unittest.TestCase):
         return model
 
     def get_generator(self, seed=0):
-        generator_device = "cpu" if not torch_device.startswith("cuda") else "cuda"
-        if torch_device != "mps":
-            return torch.Generator(device=generator_device).manual_seed(seed)
-        return torch.manual_seed(seed)
+        if torch_device == "mps":
+            return torch.manual_seed(seed)
+        return torch.Generator(device=torch_device).manual_seed(seed)
 
     @parameterized.expand(
         [
@@ -631,7 +623,7 @@ class AutoencoderKLIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator_with_fp16
+    @require_torch_gpu
     def test_stable_diffusion_fp16(self, seed, expected_slice):
         model = self.get_sd_vae_model(fp16=True)
         image = self.get_sd_image(seed, fp16=True)
@@ -685,8 +677,7 @@ class AutoencoderKLIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator
-    @skip_mps
+    @require_torch_gpu
     def test_stable_diffusion_decode(self, seed, expected_slice):
         model = self.get_sd_vae_model()
         encoding = self.get_sd_image(seed, shape=(3, 4, 64, 64))
@@ -709,7 +700,7 @@ class AutoencoderKLIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator_with_fp16
+    @require_torch_gpu
     def test_stable_diffusion_decode_fp16(self, seed, expected_slice):
         model = self.get_sd_vae_model(fp16=True)
         encoding = self.get_sd_image(seed, shape=(3, 4, 64, 64), fp16=True)
@@ -820,7 +811,7 @@ class AsymmetricAutoencoderKLIntegrationTests(unittest.TestCase):
         # clean up the VRAM after each test
         super().tearDown()
         gc.collect()
-        backend_empty_cache(torch_device)
+        torch.cuda.empty_cache()
 
     def get_sd_image(self, seed=0, shape=(4, 3, 512, 512), fp16=False):
         dtype = torch.float16 if fp16 else torch.float32
@@ -841,10 +832,9 @@ class AsymmetricAutoencoderKLIntegrationTests(unittest.TestCase):
         return model
 
     def get_generator(self, seed=0):
-        generator_device = "cpu" if not torch_device.startswith("cuda") else "cuda"
-        if torch_device != "mps":
-            return torch.Generator(device=generator_device).manual_seed(seed)
-        return torch.manual_seed(seed)
+        if torch_device == "mps":
+            return torch.manual_seed(seed)
+        return torch.Generator(device=torch_device).manual_seed(seed)
 
     @parameterized.expand(
         [
@@ -915,8 +905,7 @@ class AsymmetricAutoencoderKLIntegrationTests(unittest.TestCase):
             # fmt: on
         ]
     )
-    @require_torch_accelerator
-    @skip_mps
+    @require_torch_gpu
     def test_stable_diffusion_decode(self, seed, expected_slice):
         model = self.get_sd_vae_model()
         encoding = self.get_sd_image(seed, shape=(3, 4, 64, 64))

tests/models/test_models_vq.py

@@ -18,12 +18,7 @@ import unittest
 import torch
 
 from diffusers import VQModel
-from diffusers.utils.testing_utils import (
-    backend_manual_seed,
-    enable_full_determinism,
-    floats_tensor,
-    torch_device,
-)
+from diffusers.utils.testing_utils import enable_full_determinism, floats_tensor, torch_device
 
 from .test_modeling_common import ModelTesterMixin, UNetTesterMixin
 
@@ -85,7 +80,8 @@ class VQModelTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
         model.to(torch_device).eval()
 
         torch.manual_seed(0)
-        backend_manual_seed(torch_device, 0)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
         image = torch.randn(1, model.config.in_channels, model.config.sample_size, model.config.sample_size)
         image = image.to(torch_device)

tests/models/test_unet_blocks_common.py

@@ -12,17 +12,12 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import unittest
 from typing import Tuple
 
 import torch
 
-from diffusers.utils.testing_utils import (
-    floats_tensor,
-    require_torch,
-    require_torch_accelerator_with_training,
-    torch_all_close,
-    torch_device,
-)
+from diffusers.utils.testing_utils import floats_tensor, require_torch, torch_all_close, torch_device
 from diffusers.utils.torch_utils import randn_tensor
 
 
@@ -109,7 +104,7 @@ class UNetBlockTesterMixin:
         expected_slice = torch.tensor(expected_slice).to(torch_device)
         assert torch_all_close(output_slice.flatten(), expected_slice, atol=5e-3)
 
-    @require_torch_accelerator_with_training
+    @unittest.skipIf(torch_device == "mps", "Training is not supported in mps")
     def test_training(self):
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
         model = self.block_class(**init_dict)

tests/pipelines/controlnetxs/test_controlnetxs.py

@@ -1,306 +0,0 @@
-# coding=utf-8
-# Copyright 2023 HuggingFace Inc.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import gc
-import traceback
-import unittest
-
-import numpy as np
-import torch
-from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
-
-from diffusers import (
-    AutoencoderKL,
-    ControlNetXSModel,
-    DDIMScheduler,
-    LCMScheduler,
-    StableDiffusionControlNetXSPipeline,
-    UNet2DConditionModel,
-)
-from diffusers.utils.import_utils import is_xformers_available
-from diffusers.utils.testing_utils import (
-    enable_full_determinism,
-    load_image,
-    load_numpy,
-    require_python39_or_higher,
-    require_torch_2,
-    require_torch_gpu,
-    run_test_in_subprocess,
-    slow,
-    torch_device,
-)
-from diffusers.utils.torch_utils import randn_tensor
-
-from ..pipeline_params import (
-    IMAGE_TO_IMAGE_IMAGE_PARAMS,
-    TEXT_TO_IMAGE_BATCH_PARAMS,
-    TEXT_TO_IMAGE_IMAGE_PARAMS,
-    TEXT_TO_IMAGE_PARAMS,
-)
-from ..test_pipelines_common import (
-    PipelineKarrasSchedulerTesterMixin,
-    PipelineLatentTesterMixin,
-    PipelineTesterMixin,
-)
-
-
-enable_full_determinism()
-
-
-# Will be run via run_test_in_subprocess
-def _test_stable_diffusion_compile(in_queue, out_queue, timeout):
-    error = None
-    try:
-        _ = in_queue.get(timeout=timeout)
-
-        controlnet = ControlNetXSModel.from_pretrained("UmerHA/ConrolNetXS-SD2.1-canny")
-
-        pipe = StableDiffusionControlNetXSPipeline.from_pretrained(
-            "stabilityai/stable-diffusion-2-1", safety_checker=None, controlnet=controlnet
-        )
-        pipe.to("cuda")
-        pipe.set_progress_bar_config(disable=None)
-
-        pipe.unet.to(memory_format=torch.channels_last)
-        pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
-
-        pipe.controlnet.to(memory_format=torch.channels_last)
-        pipe.controlnet = torch.compile(pipe.controlnet, mode="reduce-overhead", fullgraph=True)
-
-        generator = torch.Generator(device="cpu").manual_seed(0)
-        prompt = "bird"
-        image = load_image(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/bird_canny.png"
-        ).resize((512, 512))
-
-        output = pipe(prompt, image, num_inference_steps=10, generator=generator, output_type="np")
-        image = output.images[0]
-
-        assert image.shape == (512, 512, 3)
-
-        expected_image = load_numpy(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/bird_canny_out_full.npy"
-        )
-        expected_image = np.resize(expected_image, (512, 512, 3))
-
-        assert np.abs(expected_image - image).max() < 1.0
-
-    except Exception:
-        error = f"{traceback.format_exc()}"
-
-    results = {"error": error}
-    out_queue.put(results, timeout=timeout)
-    out_queue.join()
-
-
-class ControlNetXSPipelineFastTests(
-    PipelineLatentTesterMixin, PipelineKarrasSchedulerTesterMixin, PipelineTesterMixin, unittest.TestCase
-):
-    pipeline_class = StableDiffusionControlNetXSPipeline
-    params = TEXT_TO_IMAGE_PARAMS
-    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
-    image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
-    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
-
-    def get_dummy_components(self, time_cond_proj_dim=None):
-        torch.manual_seed(0)
-        unet = UNet2DConditionModel(
-            block_out_channels=(4, 8),
-            layers_per_block=2,
-            sample_size=32,
-            in_channels=4,
-            out_channels=4,
-            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
-            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
-            cross_attention_dim=32,
-            norm_num_groups=1,
-            time_cond_proj_dim=time_cond_proj_dim,
-        )
-        torch.manual_seed(0)
-        controlnet = ControlNetXSModel.from_unet(
-            unet=unet,
-            time_embedding_mix=0.95,
-            learn_embedding=True,
-            size_ratio=0.5,
-            conditioning_embedding_out_channels=(16, 32),
-            num_attention_heads=2,
-        )
-        torch.manual_seed(0)
-        scheduler = DDIMScheduler(
-            beta_start=0.00085,
-            beta_end=0.012,
-            beta_schedule="scaled_linear",
-            clip_sample=False,
-            set_alpha_to_one=False,
-        )
-        torch.manual_seed(0)
-        vae = AutoencoderKL(
-            block_out_channels=[4, 8],
-            in_channels=3,
-            out_channels=3,
-            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
-            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
-            latent_channels=4,
-            norm_num_groups=2,
-        )
-        torch.manual_seed(0)
-        text_encoder_config = CLIPTextConfig(
-            bos_token_id=0,
-            eos_token_id=2,
-            hidden_size=32,
-            intermediate_size=37,
-            layer_norm_eps=1e-05,
-            num_attention_heads=4,
-            num_hidden_layers=5,
-            pad_token_id=1,
-            vocab_size=1000,
-        )
-        text_encoder = CLIPTextModel(text_encoder_config)
-        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
-
-        components = {
-            "unet": unet,
-            "controlnet": controlnet,
-            "scheduler": scheduler,
-            "vae": vae,
-            "text_encoder": text_encoder,
-            "tokenizer": tokenizer,
-            "safety_checker": None,
-            "feature_extractor": None,
-        }
-        return components
-
-    def get_dummy_inputs(self, device, seed=0):
-        if str(device).startswith("mps"):
-            generator = torch.manual_seed(seed)
-        else:
-            generator = torch.Generator(device=device).manual_seed(seed)
-
-        controlnet_embedder_scale_factor = 2
-        image = randn_tensor(
-            (1, 3, 32 * controlnet_embedder_scale_factor, 32 * controlnet_embedder_scale_factor),
-            generator=generator,
-            device=torch.device(device),
-        )
-
-        inputs = {
-            "prompt": "A painting of a squirrel eating a burger",
-            "generator": generator,
-            "num_inference_steps": 2,
-            "guidance_scale": 6.0,
-            "output_type": "numpy",
-            "image": image,
-        }
-
-        return inputs
-
-    def test_attention_slicing_forward_pass(self):
-        return self._test_attention_slicing_forward_pass(expected_max_diff=2e-3)
-
-    @unittest.skipIf(
-        torch_device != "cuda" or not is_xformers_available(),
-        reason="XFormers attention is only available with CUDA and `xformers` installed",
-    )
-    def test_xformers_attention_forwardGenerator_pass(self):
-        self._test_xformers_attention_forwardGenerator_pass(expected_max_diff=2e-3)
-
-    def test_inference_batch_single_identical(self):
-        self._test_inference_batch_single_identical(expected_max_diff=2e-3)
-
-    def test_controlnet_lcm(self):
-        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
-
-        components = self.get_dummy_components(time_cond_proj_dim=256)
-        sd_pipe = StableDiffusionControlNetXSPipeline(**components)
-        sd_pipe.scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
-        sd_pipe = sd_pipe.to(torch_device)
-        sd_pipe.set_progress_bar_config(disable=None)
-
-        inputs = self.get_dummy_inputs(device)
-        output = sd_pipe(**inputs)
-        image = output.images
-
-        image_slice = image[0, -3:, -3:, -1]
-
-        assert image.shape == (1, 64, 64, 3)
-        expected_slice = np.array(
-            [0.52700454, 0.3930534, 0.25509018, 0.7132304, 0.53696585, 0.46568912, 0.7095368, 0.7059624, 0.4744786]
-        )
-
-        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
-
-
-@slow
-@require_torch_gpu
-class ControlNetXSPipelineSlowTests(unittest.TestCase):
-    def tearDown(self):
-        super().tearDown()
-        gc.collect()
-        torch.cuda.empty_cache()
-
-    def test_canny(self):
-        controlnet = ControlNetXSModel.from_pretrained("UmerHA/ConrolNetXS-SD2.1-canny")
-
-        pipe = StableDiffusionControlNetXSPipeline.from_pretrained(
-            "stabilityai/stable-diffusion-2-1", safety_checker=None, controlnet=controlnet
-        )
-        pipe.enable_model_cpu_offload()
-        pipe.set_progress_bar_config(disable=None)
-
-        generator = torch.Generator(device="cpu").manual_seed(0)
-        prompt = "bird"
-        image = load_image(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/bird_canny.png"
-        )
-
-        output = pipe(prompt, image, generator=generator, output_type="np", num_inference_steps=3)
-
-        image = output.images[0]
-
-        assert image.shape == (768, 512, 3)
-
-        original_image = image[-3:, -3:, -1].flatten()
-        expected_image = np.array([0.1274, 0.1401, 0.147, 0.1185, 0.1555, 0.1492, 0.1565, 0.1474, 0.1701])
-        assert np.allclose(original_image, expected_image, atol=1e-04)
-
-    def test_depth(self):
-        controlnet = ControlNetXSModel.from_pretrained("UmerHA/ConrolNetXS-SD2.1-depth")
-
-        pipe = StableDiffusionControlNetXSPipeline.from_pretrained(
-            "stabilityai/stable-diffusion-2-1", safety_checker=None, controlnet=controlnet
-        )
-        pipe.enable_model_cpu_offload()
-        pipe.set_progress_bar_config(disable=None)
-
-        generator = torch.Generator(device="cpu").manual_seed(0)
-        prompt = "Stormtrooper's lecture"
-        image = load_image(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/stormtrooper_depth.png"
-        )
-
-        output = pipe(prompt, image, generator=generator, output_type="np", num_inference_steps=3)
-
-        image = output.images[0]
-
-        assert image.shape == (512, 512, 3)
-
-        original_image = image[-3:, -3:, -1].flatten()
-        expected_image = np.array([0.1098, 0.1025, 0.1211, 0.1129, 0.1165, 0.1262, 0.1185, 0.1261, 0.1703])
-        assert np.allclose(original_image, expected_image, atol=1e-04)
-
-    @require_python39_or_higher
-    @require_torch_2
-    def test_stable_diffusion_compile(self):
-        run_test_in_subprocess(test_case=self, target_func=_test_stable_diffusion_compile, inputs=None)

tests/pipelines/controlnetxs/test_controlnetxs_sdxl.py

@@ -1,362 +0,0 @@
-# coding=utf-8
-# Copyright 2023 HuggingFace Inc.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import gc
-import unittest
-
-import numpy as np
-import torch
-from transformers import CLIPTextConfig, CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
-
-from diffusers import (
-    AutoencoderKL,
-    ControlNetXSModel,
-    EulerDiscreteScheduler,
-    StableDiffusionXLControlNetXSPipeline,
-    UNet2DConditionModel,
-)
-from diffusers.utils.import_utils import is_xformers_available
-from diffusers.utils.testing_utils import enable_full_determinism, load_image, require_torch_gpu, slow, torch_device
-from diffusers.utils.torch_utils import randn_tensor
-
-from ..pipeline_params import (
-    IMAGE_TO_IMAGE_IMAGE_PARAMS,
-    TEXT_TO_IMAGE_BATCH_PARAMS,
-    TEXT_TO_IMAGE_IMAGE_PARAMS,
-    TEXT_TO_IMAGE_PARAMS,
-)
-from ..test_pipelines_common import (
-    PipelineKarrasSchedulerTesterMixin,
-    PipelineLatentTesterMixin,
-    PipelineTesterMixin,
-    SDXLOptionalComponentsTesterMixin,
-)
-
-
-enable_full_determinism()
-
-
-class StableDiffusionXLControlNetXSPipelineFastTests(
-    PipelineLatentTesterMixin,
-    PipelineKarrasSchedulerTesterMixin,
-    PipelineTesterMixin,
-    SDXLOptionalComponentsTesterMixin,
-    unittest.TestCase,
-):
-    pipeline_class = StableDiffusionXLControlNetXSPipeline
-    params = TEXT_TO_IMAGE_PARAMS
-    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
-    image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
-    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
-
-    def get_dummy_components(self):
-        torch.manual_seed(0)
-        unet = UNet2DConditionModel(
-            block_out_channels=(32, 64),
-            layers_per_block=2,
-            sample_size=32,
-            in_channels=4,
-            out_channels=4,
-            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
-            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
-            # SD2-specific config below
-            attention_head_dim=(2, 4),
-            use_linear_projection=True,
-            addition_embed_type="text_time",
-            addition_time_embed_dim=8,
-            transformer_layers_per_block=(1, 2),
-            projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
-            cross_attention_dim=64,
-        )
-        torch.manual_seed(0)
-        controlnet = ControlNetXSModel.from_unet(
-            unet,
-            time_embedding_mix=0.95,
-            learn_embedding=True,
-            size_ratio=0.5,
-            conditioning_embedding_out_channels=(16, 32),
-        )
-        torch.manual_seed(0)
-        scheduler = EulerDiscreteScheduler(
-            beta_start=0.00085,
-            beta_end=0.012,
-            steps_offset=1,
-            beta_schedule="scaled_linear",
-            timestep_spacing="leading",
-        )
-        torch.manual_seed(0)
-        vae = AutoencoderKL(
-            block_out_channels=[32, 64],
-            in_channels=3,
-            out_channels=3,
-            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
-            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
-            latent_channels=4,
-        )
-        torch.manual_seed(0)
-        text_encoder_config = CLIPTextConfig(
-            bos_token_id=0,
-            eos_token_id=2,
-            hidden_size=32,
-            intermediate_size=37,
-            layer_norm_eps=1e-05,
-            num_attention_heads=4,
-            num_hidden_layers=5,
-            pad_token_id=1,
-            vocab_size=1000,
-            # SD2-specific config below
-            hidden_act="gelu",
-            projection_dim=32,
-        )
-        text_encoder = CLIPTextModel(text_encoder_config)
-        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
-
-        text_encoder_2 = CLIPTextModelWithProjection(text_encoder_config)
-        tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
-
-        components = {
-            "unet": unet,
-            "controlnet": controlnet,
-            "scheduler": scheduler,
-            "vae": vae,
-            "text_encoder": text_encoder,
-            "tokenizer": tokenizer,
-            "text_encoder_2": text_encoder_2,
-            "tokenizer_2": tokenizer_2,
-        }
-        return components
-
-    # copied from test_controlnet_sdxl.py
-    def get_dummy_inputs(self, device, seed=0):
-        if str(device).startswith("mps"):
-            generator = torch.manual_seed(seed)
-        else:
-            generator = torch.Generator(device=device).manual_seed(seed)
-
-        controlnet_embedder_scale_factor = 2
-        image = randn_tensor(
-            (1, 3, 32 * controlnet_embedder_scale_factor, 32 * controlnet_embedder_scale_factor),
-            generator=generator,
-            device=torch.device(device),
-        )
-
-        inputs = {
-            "prompt": "A painting of a squirrel eating a burger",
-            "generator": generator,
-            "num_inference_steps": 2,
-            "guidance_scale": 6.0,
-            "output_type": "np",
-            "image": image,
-        }
-
-        return inputs
-
-    # copied from test_controlnet_sdxl.py
-    def test_attention_slicing_forward_pass(self):
-        return self._test_attention_slicing_forward_pass(expected_max_diff=2e-3)
-
-    # copied from test_controlnet_sdxl.py
-    @unittest.skipIf(
-        torch_device != "cuda" or not is_xformers_available(),
-        reason="XFormers attention is only available with CUDA and `xformers` installed",
-    )
-    def test_xformers_attention_forwardGenerator_pass(self):
-        self._test_xformers_attention_forwardGenerator_pass(expected_max_diff=2e-3)
-
-    # copied from test_controlnet_sdxl.py
-    def test_inference_batch_single_identical(self):
-        self._test_inference_batch_single_identical(expected_max_diff=2e-3)
-
-    # copied from test_controlnet_sdxl.py
-    def test_save_load_optional_components(self):
-        self._test_save_load_optional_components()
-
-    # copied from test_controlnet_sdxl.py
-    @require_torch_gpu
-    def test_stable_diffusion_xl_offloads(self):
-        pipes = []
-        components = self.get_dummy_components()
-        sd_pipe = self.pipeline_class(**components).to(torch_device)
-        pipes.append(sd_pipe)
-
-        components = self.get_dummy_components()
-        sd_pipe = self.pipeline_class(**components)
-        sd_pipe.enable_model_cpu_offload()
-        pipes.append(sd_pipe)
-
-        components = self.get_dummy_components()
-        sd_pipe = self.pipeline_class(**components)
-        sd_pipe.enable_sequential_cpu_offload()
-        pipes.append(sd_pipe)
-
-        image_slices = []
-        for pipe in pipes:
-            pipe.unet.set_default_attn_processor()
-
-            inputs = self.get_dummy_inputs(torch_device)
-            image = pipe(**inputs).images
-
-            image_slices.append(image[0, -3:, -3:, -1].flatten())
-
-        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
-        assert np.abs(image_slices[0] - image_slices[2]).max() < 1e-3
-
-    # copied from test_controlnet_sdxl.py
-    def test_stable_diffusion_xl_multi_prompts(self):
-        components = self.get_dummy_components()
-        sd_pipe = self.pipeline_class(**components).to(torch_device)
-
-        # forward with single prompt
-        inputs = self.get_dummy_inputs(torch_device)
-        output = sd_pipe(**inputs)
-        image_slice_1 = output.images[0, -3:, -3:, -1]
-
-        # forward with same prompt duplicated
-        inputs = self.get_dummy_inputs(torch_device)
-        inputs["prompt_2"] = inputs["prompt"]
-        output = sd_pipe(**inputs)
-        image_slice_2 = output.images[0, -3:, -3:, -1]
-
-        # ensure the results are equal
-        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4
-
-        # forward with different prompt
-        inputs = self.get_dummy_inputs(torch_device)
-        inputs["prompt_2"] = "different prompt"
-        output = sd_pipe(**inputs)
-        image_slice_3 = output.images[0, -3:, -3:, -1]
-
-        # ensure the results are not equal
-        assert np.abs(image_slice_1.flatten() - image_slice_3.flatten()).max() > 1e-4
-
-        # manually set a negative_prompt
-        inputs = self.get_dummy_inputs(torch_device)
-        inputs["negative_prompt"] = "negative prompt"
-        output = sd_pipe(**inputs)
-        image_slice_1 = output.images[0, -3:, -3:, -1]
-
-        # forward with same negative_prompt duplicated
-        inputs = self.get_dummy_inputs(torch_device)
-        inputs["negative_prompt"] = "negative prompt"
-        inputs["negative_prompt_2"] = inputs["negative_prompt"]
-        output = sd_pipe(**inputs)
-        image_slice_2 = output.images[0, -3:, -3:, -1]
-
-        # ensure the results are equal
-        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4
-
-        # forward with different negative_prompt
-        inputs = self.get_dummy_inputs(torch_device)
-        inputs["negative_prompt"] = "negative prompt"
-        inputs["negative_prompt_2"] = "different negative prompt"
-        output = sd_pipe(**inputs)
-        image_slice_3 = output.images[0, -3:, -3:, -1]
-
-        # ensure the results are not equal
-        assert np.abs(image_slice_1.flatten() - image_slice_3.flatten()).max() > 1e-4
-
-    # copied from test_stable_diffusion_xl.py
-    def test_stable_diffusion_xl_prompt_embeds(self):
-        components = self.get_dummy_components()
-        sd_pipe = self.pipeline_class(**components)
-        sd_pipe = sd_pipe.to(torch_device)
-        sd_pipe = sd_pipe.to(torch_device)
-        sd_pipe.set_progress_bar_config(disable=None)
-
-        # forward without prompt embeds
-        inputs = self.get_dummy_inputs(torch_device)
-        inputs["prompt"] = 2 * [inputs["prompt"]]
-        inputs["num_images_per_prompt"] = 2
-
-        output = sd_pipe(**inputs)
-        image_slice_1 = output.images[0, -3:, -3:, -1]
-
-        # forward with prompt embeds
-        inputs = self.get_dummy_inputs(torch_device)
-        prompt = 2 * [inputs.pop("prompt")]
-
-        (
-            prompt_embeds,
-            negative_prompt_embeds,
-            pooled_prompt_embeds,
-            negative_pooled_prompt_embeds,
-        ) = sd_pipe.encode_prompt(prompt)
-
-        output = sd_pipe(
-            **inputs,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            pooled_prompt_embeds=pooled_prompt_embeds,
-            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
-        )
-        image_slice_2 = output.images[0, -3:, -3:, -1]
-
-        # make sure that it's equal
-        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1.1e-4
-
-
-@slow
-@require_torch_gpu
-class ControlNetSDXLPipelineXSSlowTests(unittest.TestCase):
-    def tearDown(self):
-        super().tearDown()
-        gc.collect()
-        torch.cuda.empty_cache()
-
-    def test_canny(self):
-        controlnet = ControlNetXSModel.from_pretrained("UmerHA/ConrolNetXS-SDXL-canny")
-
-        pipe = StableDiffusionXLControlNetXSPipeline.from_pretrained(
-            "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet
-        )
-        pipe.enable_sequential_cpu_offload()
-        pipe.set_progress_bar_config(disable=None)
-
-        generator = torch.Generator(device="cpu").manual_seed(0)
-        prompt = "bird"
-        image = load_image(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/bird_canny.png"
-        )
-
-        images = pipe(prompt, image=image, generator=generator, output_type="np", num_inference_steps=3).images
-
-        assert images[0].shape == (768, 512, 3)
-
-        original_image = images[0, -3:, -3:, -1].flatten()
-        expected_image = np.array([0.4359, 0.4335, 0.4609, 0.4515, 0.4669, 0.4494, 0.452, 0.4493, 0.4382])
-        assert np.allclose(original_image, expected_image, atol=1e-04)
-
-    def test_depth(self):
-        controlnet = ControlNetXSModel.from_pretrained("UmerHA/ConrolNetXS-SDXL-depth")
-
-        pipe = StableDiffusionXLControlNetXSPipeline.from_pretrained(
-            "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet
-        )
-        pipe.enable_sequential_cpu_offload()
-        pipe.set_progress_bar_config(disable=None)
-
-        generator = torch.Generator(device="cpu").manual_seed(0)
-        prompt = "Stormtrooper's lecture"
-        image = load_image(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/stormtrooper_depth.png"
-        )
-
-        images = pipe(prompt, image=image, generator=generator, output_type="np", num_inference_steps=3).images
-
-        assert images[0].shape == (512, 512, 3)
-
-        original_image = images[0, -3:, -3:, -1].flatten()
-        expected_image = np.array([0.4411, 0.3617, 0.2654, 0.266, 0.3449, 0.3898, 0.3745, 0.353, 0.326])
-        assert np.allclose(original_image, expected_image, atol=1e-04)

tests/pipelines/ip_adapters/test_ip_adapter_stable_diffusion.py

@@ -182,25 +182,6 @@ class IPAdapterSDIntegrationTests(IPAdapterNightlyTestsMixin):
 
         assert np.allclose(image_slice, expected_slice, atol=1e-4, rtol=1e-4)
 
-    def test_text_to_image_full_face(self):
-        image_encoder = self.get_image_encoder(repo_id="h94/IP-Adapter", subfolder="models/image_encoder")
-        pipeline = StableDiffusionPipeline.from_pretrained(
-            "runwayml/stable-diffusion-v1-5", image_encoder=image_encoder, safety_checker=None, torch_dtype=self.dtype
-        )
-        pipeline.to(torch_device)
-        pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter-full-face_sd15.bin")
-        pipeline.set_ip_adapter_scale(0.7)
-
-        inputs = self.get_dummy_inputs()
-        images = pipeline(**inputs).images
-        image_slice = images[0, :3, :3, -1].flatten()
-
-        expected_slice = np.array(
-            [0.1706543, 0.1303711, 0.12573242, 0.21777344, 0.14550781, 0.14038086, 0.40820312, 0.41455078, 0.42529297]
-        )
-
-        assert np.allclose(image_slice, expected_slice, atol=1e-4, rtol=1e-4)
-
 
 @slow
 @require_torch_gpu

tests/pipelines/stable_diffusion_2/test_stable_diffusion.py

@@ -34,14 +34,11 @@ from diffusers import (
 )
 from diffusers.utils.testing_utils import (
     CaptureLogger,
-    backend_empty_cache,
     enable_full_determinism,
     load_numpy,
     nightly,
     numpy_cosine_similarity_distance,
-    require_torch_accelerator,
     require_torch_gpu,
-    skip_mps,
     slow,
     torch_device,
 )
@@ -131,12 +128,10 @@ class StableDiffusion2PipelineFastTests(
         return components
 
     def get_dummy_inputs(self, device, seed=0):
-        generator_device = "cpu" if not device.startswith("cuda") else "cuda"
-        if not str(device).startswith("mps"):
-            generator = torch.Generator(device=generator_device).manual_seed(seed)
-        else:
+        if str(device).startswith("mps"):
             generator = torch.manual_seed(seed)
-
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
         inputs = {
             "prompt": "A painting of a squirrel eating a burger",
             "generator": generator,
@@ -304,21 +299,15 @@ class StableDiffusion2PipelineFastTests(
 
 
 @slow
-@require_torch_accelerator
-@skip_mps
+@require_torch_gpu
 class StableDiffusion2PipelineSlowTests(unittest.TestCase):
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        backend_empty_cache(torch_device)
+        torch.cuda.empty_cache()
 
     def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
-        _generator_device = "cpu" if not generator_device.startswith("cuda") else "cuda"
-        if not str(device).startswith("mps"):
-            generator = torch.Generator(device=_generator_device).manual_seed(seed)
-        else:
-            generator = torch.manual_seed(seed)
-
+        generator = torch.Generator(device=generator_device).manual_seed(seed)
         latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
         latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
         inputs = {
@@ -372,7 +361,6 @@ class StableDiffusion2PipelineSlowTests(unittest.TestCase):
         expected_slice = np.array([0.10440, 0.13115, 0.11100, 0.10141, 0.11440, 0.07215, 0.11332, 0.09693, 0.10006])
         assert np.abs(image_slice - expected_slice).max() < 3e-3
 
-    @require_torch_gpu
     def test_stable_diffusion_attention_slicing(self):
         torch.cuda.reset_peak_memory_stats()
         pipe = StableDiffusionPipeline.from_pretrained(
@@ -444,7 +432,6 @@ class StableDiffusion2PipelineSlowTests(unittest.TestCase):
         assert callback_fn.has_been_called
         assert number_of_steps == inputs["num_inference_steps"]
 
-    @require_torch_gpu
     def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self):
         torch.cuda.empty_cache()
         torch.cuda.reset_max_memory_allocated()
@@ -465,7 +452,6 @@ class StableDiffusion2PipelineSlowTests(unittest.TestCase):
         # make sure that less than 2.8 GB is allocated
         assert mem_bytes < 2.8 * 10**9
 
-    @require_torch_gpu
     def test_stable_diffusion_pipeline_with_model_offloading(self):
         torch.cuda.empty_cache()
         torch.cuda.reset_max_memory_allocated()
@@ -525,21 +511,15 @@ class StableDiffusion2PipelineSlowTests(unittest.TestCase):
 
 
 @nightly
-@require_torch_accelerator
-@skip_mps
+@require_torch_gpu
 class StableDiffusion2PipelineNightlyTests(unittest.TestCase):
     def tearDown(self):
         super().tearDown()
         gc.collect()
-        backend_empty_cache(torch_device)
+        torch.cuda.empty_cache()
 
     def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
-        _generator_device = "cpu" if not generator_device.startswith("cuda") else "cuda"
-        if not str(device).startswith("mps"):
-            generator = torch.Generator(device=_generator_device).manual_seed(seed)
-        else:
-            generator = torch.manual_seed(seed)
-
+        generator = torch.Generator(device=generator_device).manual_seed(seed)
         latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
         latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
         inputs = {

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl.py

@@ -938,37 +938,6 @@ class StableDiffusionXLPipelineFastTests(
 
         assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
 
-    def test_stable_diffusion_xl_with_fused_qkv_projections(self):
-        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
-        components = self.get_dummy_components()
-        sd_pipe = StableDiffusionXLPipeline(**components)
-        sd_pipe = sd_pipe.to(device)
-        sd_pipe.set_progress_bar_config(disable=None)
-
-        inputs = self.get_dummy_inputs(device)
-        image = sd_pipe(**inputs).images
-        original_image_slice = image[0, -3:, -3:, -1]
-
-        sd_pipe.fuse_qkv_projections()
-        inputs = self.get_dummy_inputs(device)
-        image = sd_pipe(**inputs).images
-        image_slice_fused = image[0, -3:, -3:, -1]
-
-        sd_pipe.unfuse_qkv_projections()
-        inputs = self.get_dummy_inputs(device)
-        image = sd_pipe(**inputs).images
-        image_slice_disabled = image[0, -3:, -3:, -1]
-
-        assert np.allclose(
-            original_image_slice, image_slice_fused, atol=1e-2, rtol=1e-2
-        ), "Fusion of QKV projections shouldn't affect the outputs."
-        assert np.allclose(
-            image_slice_fused, image_slice_disabled, atol=1e-2, rtol=1e-2
-        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
-        assert np.allclose(
-            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
-        ), "Original outputs should match when fused QKV projections are disabled."
-
 
 @slow
 class StableDiffusionXLPipelineIntegrationTests(unittest.TestCase):

tests/schedulers/test_scheduler_euler.py

@@ -45,10 +45,6 @@ class EulerDiscreteSchedulerTest(SchedulerCommonTest):
     def test_karras_sigmas(self):
         self.check_over_configs(use_karras_sigmas=True, sigma_min=0.02, sigma_max=700.0)
 
-    def test_rescale_betas_zero_snr(self):
-        for rescale_betas_zero_snr in [True, False]:
-            self.check_over_configs(rescale_betas_zero_snr=rescale_betas_zero_snr)
-
     def test_full_loop_no_noise(self):
         scheduler_class = self.scheduler_classes[0]
         scheduler_config = self.get_scheduler_config()