Merge branch 'main' into model-test-refactor

CONTRIBUTING.md

@@ -1 +0,0 @@
-docs/source/en/conceptual/contribution.md

CONTRIBUTING.md

@@ -0,0 +1,506 @@
+<!--Copyright 2025 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.
+-->
+
+# How to contribute to Diffusers 🧨
+
+We ❤️ contributions from the open-source community! Everyone is welcome, and all types of participation –not just code– are valued and appreciated. Answering questions, helping others, reaching out, and improving the documentation are all immensely valuable to the community, so don't be afraid and get involved if you're up for it!
+
+Everyone is encouraged to start by saying 👋 in our public Discord channel. We discuss the latest trends in diffusion models, ask questions, show off personal projects, help each other with contributions, or just hang out ☕. <a href="https://discord.gg/G7tWnz98XR"><img alt="Join us on Discord" src="https://img.shields.io/discord/823813159592001537?color=5865F2&logo=Discord&logoColor=white"></a>
+
+Whichever way you choose to contribute, we strive to be part of an open, welcoming, and kind community. Please, read our [code of conduct](https://github.com/huggingface/diffusers/blob/main/CODE_OF_CONDUCT.md) and be mindful to respect it during your interactions. We also recommend you become familiar with the [ethical guidelines](https://huggingface.co/docs/diffusers/conceptual/ethical_guidelines) that guide our project and ask you to adhere to the same principles of transparency and responsibility.
+
+We enormously value feedback from the community, so please do not be afraid to speak up if you believe you have valuable feedback that can help improve the library - every message, comment, issue, and pull request (PR) is read and considered.
+
+## Overview
+
+You can contribute in many ways ranging from answering questions on issues to adding new diffusion models to
+the core library.
+
+In the following, we give an overview of different ways to contribute, ranked by difficulty in ascending order. All of them are valuable to the community.
+
+* 1. Asking and answering questions on [the Diffusers discussion forum](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers) or on [Discord](https://discord.gg/G7tWnz98XR).
+* 2. Opening new issues on [the GitHub Issues tab](https://github.com/huggingface/diffusers/issues/new/choose).
+* 3. Answering issues on [the GitHub Issues tab](https://github.com/huggingface/diffusers/issues).
+* 4. Fix a simple issue, marked by the "Good first issue" label, see [here](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22good+first+issue%22).
+* 5. Contribute to the [documentation](https://github.com/huggingface/diffusers/tree/main/docs/source).
+* 6. Contribute a [Community Pipeline](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3Acommunity-examples).
+* 7. Contribute to the [examples](https://github.com/huggingface/diffusers/tree/main/examples).
+* 8. Fix a more difficult issue, marked by the "Good second issue" label, see [here](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22Good+second+issue%22).
+* 9. Add a new pipeline, model, or scheduler, see ["New Pipeline/Model"](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22New+pipeline%2Fmodel%22) and ["New scheduler"](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22New+scheduler%22) issues. For this contribution, please have a look at [Design Philosophy](https://github.com/huggingface/diffusers/blob/main/PHILOSOPHY.md).
+
+As said before, **all contributions are valuable to the community**.
+In the following, we will explain each contribution a bit more in detail.
+
+For all contributions 4-9, you will need to open a PR. It is explained in detail how to do so in [Opening a pull request](#how-to-open-a-pr).
+
+### 1. Asking and answering questions on the Diffusers discussion forum or on the Diffusers Discord
+
+Any question or comment related to the Diffusers library can be asked on the [discussion forum](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/) or on [Discord](https://discord.gg/G7tWnz98XR). Such questions and comments include (but are not limited to):
+- Reports of training or inference experiments in an attempt to share knowledge
+- Presentation of personal projects
+- Questions to non-official training examples
+- Project proposals
+- General feedback
+- Paper summaries
+- Asking for help on personal projects that build on top of the Diffusers library
+- General questions
+- Ethical questions regarding diffusion models
+- ...
+
+Every question that is asked on the forum or on Discord actively encourages the community to publicly
+share knowledge and might very well help a beginner in the future who has the same question you're
+having. Please do pose any questions you might have.
+In the same spirit, you are of immense help to the community by answering such questions because this way you are publicly documenting knowledge for everybody to learn from.
+
+**Please** keep in mind that the more effort you put into asking or answering a question, the higher
+the quality of the publicly documented knowledge. In the same way, well-posed and well-answered questions create a high-quality knowledge database accessible to everybody, while badly posed questions or answers reduce the overall quality of the public knowledge database.
+In short, a high quality question or answer is *precise*, *concise*, *relevant*, *easy-to-understand*, *accessible*, and *well-formatted/well-posed*. For more information, please have a look through the [How to write a good issue](#how-to-write-a-good-issue) section.
+
+**NOTE about channels**:
+[*The forum*](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63) is much better indexed by search engines, such as Google. Posts are ranked by popularity rather than chronologically. Hence, it's easier to look up questions and answers that we posted some time ago.
+In addition, questions and answers posted in the forum can easily be linked to.
+In contrast, *Discord* has a chat-like format that invites fast back-and-forth communication.
+While it will most likely take less time for you to get an answer to your question on Discord, your
+question won't be visible anymore over time. Also, it's much harder to find information that was posted a while back on Discord. We therefore strongly recommend using the forum for high-quality questions and answers in an attempt to create long-lasting knowledge for the community. If discussions on Discord lead to very interesting answers and conclusions, we recommend posting the results on the forum to make the information more available for future readers.
+
+### 2. Opening new issues on the GitHub issues tab
+
+The 🧨 Diffusers library is robust and reliable thanks to the users who notify us of
+the problems they encounter. So thank you for reporting an issue.
+
+Remember, GitHub issues are reserved for technical questions directly related to the Diffusers library, bug reports, feature requests, or feedback on the library design.
+
+In a nutshell, this means that everything that is **not** related to the **code of the Diffusers library** (including the documentation) should **not** be asked on GitHub, but rather on either the [forum](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63) or [Discord](https://discord.gg/G7tWnz98XR).
+
+**Please consider the following guidelines when opening a new issue**:
+- Make sure you have searched whether your issue has already been asked before (use the search bar on GitHub under Issues).
+- Please never report a new issue on another (related) issue. If another issue is highly related, please
+open a new issue nevertheless and link to the related issue.
+- Make sure your issue is written in English. Please use one of the great, free online translation services, such as [DeepL](https://www.deepl.com/translator) to translate from your native language to English if you are not comfortable in English.
+- Check whether your issue might be solved by updating to the newest Diffusers version. Before posting your issue, please make sure that `python -c "import diffusers; print(diffusers.__version__)"` is higher or matches the latest Diffusers version.
+- Remember that the more effort you put into opening a new issue, the higher the quality of your answer will be and the better the overall quality of the Diffusers issues.
+
+New issues usually include the following.
+
+#### 2.1. Reproducible, minimal bug reports
+
+A bug report should always have a reproducible code snippet and be as minimal and concise as possible.
+This means in more detail:
+- Narrow the bug down as much as you can, **do not just dump your whole code file**.
+- Format your code.
+- Do not include any external libraries except for Diffusers depending on them.
+- **Always** provide all necessary information about your environment; for this, you can run: `diffusers-cli env` in your shell and copy-paste the displayed information to the issue.
+- Explain the issue. If the reader doesn't know what the issue is and why it is an issue, she cannot solve it.
+- **Always** make sure the reader can reproduce your issue with as little effort as possible. If your code snippet cannot be run because of missing libraries or undefined variables, the reader cannot help you. Make sure your reproducible code snippet is as minimal as possible and can be copy-pasted into a simple Python shell.
+- If in order to reproduce your issue a model and/or dataset is required, make sure the reader has access to that model or dataset. You can always upload your model or dataset to the [Hub](https://huggingface.co) to make it easily downloadable. Try to keep your model and dataset as small as possible, to make the reproduction of your issue as effortless as possible.
+
+For more information, please have a look through the [How to write a good issue](#how-to-write-a-good-issue) section.
+
+You can open a bug report [here](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=bug&projects=&template=bug-report.yml).
+
+#### 2.2. Feature requests
+
+A world-class feature request addresses the following points:
+
+1. Motivation first:
+* Is it related to a problem/frustration with the library? If so, please explain
+why. Providing a code snippet that demonstrates the problem is best.
+* Is it related to something you would need for a project? We'd love to hear
+about it!
+* Is it something you worked on and think could benefit the community?
+Awesome! Tell us what problem it solved for you.
+2. Write a *full paragraph* describing the feature;
+3. Provide a **code snippet** that demonstrates its future use;
+4. In case this is related to a paper, please attach a link;
+5. Attach any additional information (drawings, screenshots, etc.) you think may help.
+
+You can open a feature request [here](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=&template=feature_request.md&title=).
+
+#### 2.3 Feedback
+
+Feedback about the library design and why it is good or not good helps the core maintainers immensely to build a user-friendly library. To understand the philosophy behind the current design philosophy, please have a look [here](https://huggingface.co/docs/diffusers/conceptual/philosophy). If you feel like a certain design choice does not fit with the current design philosophy, please explain why and how it should be changed. If a certain design choice follows the design philosophy too much, hence restricting use cases, explain why and how it should be changed.
+If a certain design choice is very useful for you, please also leave a note as this is great feedback for future design decisions.
+
+You can open an issue about feedback [here](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=&template=feedback.md&title=).
+
+#### 2.4 Technical questions
+
+Technical questions are mainly about why certain code of the library was written in a certain way, or what a certain part of the code does. Please make sure to link to the code in question and please provide detail on
+why this part of the code is difficult to understand.
+
+You can open an issue about a technical question [here](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=bug&template=bug-report.yml).
+
+#### 2.5 Proposal to add a new model, scheduler, or pipeline
+
+If the diffusion model community released a new model, pipeline, or scheduler that you would like to see in the Diffusers library, please provide the following information:
+
+* Short description of the diffusion pipeline, model, or scheduler and link to the paper or public release.
+* Link to any of its open-source implementation.
+* Link to the model weights if they are available.
+
+If you are willing to contribute to the model yourself, let us know so we can best guide you. Also, don't forget
+to tag the original author of the component (model, scheduler, pipeline, etc.) by GitHub handle if you can find it.
+
+You can open a request for a model/pipeline/scheduler [here](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=New+model%2Fpipeline%2Fscheduler&template=new-model-addition.yml).
+
+### 3. Answering issues on the GitHub issues tab
+
+Answering issues on GitHub might require some technical knowledge of Diffusers, but we encourage everybody to give it a try even if you are not 100% certain that your answer is correct.
+Some tips to give a high-quality answer to an issue:
+- Be as concise and minimal as possible.
+- Stay on topic. An answer to the issue should concern the issue and only the issue.
+- Provide links to code, papers, or other sources that prove or encourage your point.
+- Answer in code. If a simple code snippet is the answer to the issue or shows how the issue can be solved, please provide a fully reproducible code snippet.
+
+Also, many issues tend to be simply off-topic, duplicates of other issues, or irrelevant. It is of great
+help to the maintainers if you can answer such issues, encouraging the author of the issue to be
+more precise, provide the link to a duplicated issue or redirect them to [the forum](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63) or [Discord](https://discord.gg/G7tWnz98XR).
+
+If you have verified that the issued bug report is correct and requires a correction in the source code,
+please have a look at the next sections.
+
+For all of the following contributions, you will need to open a PR. It is explained in detail how to do so in the [Opening a pull request](#how-to-open-a-pr) section.
+
+### 4. Fixing a "Good first issue"
+
+*Good first issues* are marked by the [Good first issue](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22good+first+issue%22) label. Usually, the issue already
+explains how a potential solution should look so that it is easier to fix.
+If the issue hasn't been closed and you would like to try to fix this issue, you can just leave a message "I would like to try this issue.". There are usually three scenarios:
+- a.) The issue description already proposes a fix. In this case and if the solution makes sense to you, you can open a PR or draft PR to fix it.
+- b.) The issue description does not propose a fix. In this case, you can ask what a proposed fix could look like and someone from the Diffusers team should answer shortly. If you have a good idea of how to fix it, feel free to directly open a PR.
+- c.) There is already an open PR to fix the issue, but the issue hasn't been closed yet. If the PR has gone stale, you can simply open a new PR and link to the stale PR. PRs often go stale if the original contributor who wanted to fix the issue suddenly cannot find the time anymore to proceed. This often happens in open-source and is very normal. In this case, the community will be very happy if you give it a new try and leverage the knowledge of the existing PR. If there is already a PR and it is active, you can help the author by giving suggestions, reviewing the PR or even asking whether you can contribute to the PR.
+
+
+### 5. Contribute to the documentation
+
+A good library **always** has good documentation! The official documentation is often one of the first points of contact for new users of the library, and therefore contributing to the documentation is a **highly
+valuable contribution**.
+
+Contributing to the library can have many forms:
+
+- Correcting spelling or grammatical errors.
+- Correct incorrect formatting of the docstring. If you see that the official documentation is weirdly displayed or a link is broken, we are very happy if you take some time to correct it.
+- Correct the shape or dimensions of a docstring input or output tensor.
+- Clarify documentation that is hard to understand or incorrect.
+- Update outdated code examples.
+- Translating the documentation to another language.
+
+Anything displayed on [the official Diffusers doc page](https://huggingface.co/docs/diffusers/index) is part of the official documentation and can be corrected, adjusted in the respective [documentation source](https://github.com/huggingface/diffusers/tree/main/docs/source).
+
+Please have a look at [this page](https://github.com/huggingface/diffusers/tree/main/docs) on how to verify changes made to the documentation locally.
+
+
+### 6. Contribute a community pipeline
+
+[Pipelines](https://huggingface.co/docs/diffusers/api/pipelines/overview) are usually the first point of contact between the Diffusers library and the user.
+Pipelines are examples of how to use Diffusers [models](https://huggingface.co/docs/diffusers/api/models/overview) and [schedulers](https://huggingface.co/docs/diffusers/api/schedulers/overview).
+We support two types of pipelines:
+
+- Official Pipelines
+- Community Pipelines
+
+Both official and community pipelines follow the same design and consist of the same type of components.
+
+Official pipelines are tested and maintained by the core maintainers of Diffusers. Their code
+resides in [src/diffusers/pipelines](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines).
+In contrast, community pipelines are contributed and maintained purely by the **community** and are **not** tested.
+They reside in [examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community) and while they can be accessed via the [PyPI diffusers package](https://pypi.org/project/diffusers/), their code is not part of the PyPI distribution.
+
+The reason for the distinction is that the core maintainers of the Diffusers library cannot maintain and test all
+possible ways diffusion models can be used for inference, but some of them may be of interest to the community.
+Officially released diffusion pipelines,
+such as Stable Diffusion are added to the core src/diffusers/pipelines package which ensures
+high quality of maintenance, no backward-breaking code changes, and testing.
+More bleeding edge pipelines should be added as community pipelines. If usage for a community pipeline is high, the pipeline can be moved to the official pipelines upon request from the community. This is one of the ways we strive to be a community-driven library.
+
+To add a community pipeline, one should add a <name-of-the-community>.py file to [examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community) and adapt the [examples/community/README.md](https://github.com/huggingface/diffusers/tree/main/examples/community/README.md) to include an example of the new pipeline.
+
+An example can be seen [here](https://github.com/huggingface/diffusers/pull/2400).
+
+Community pipeline PRs are only checked at a superficial level and ideally they should be maintained by their original authors.
+
+Contributing a community pipeline is a great way to understand how Diffusers models and schedulers work. Having contributed a community pipeline is usually the first stepping stone to contributing an official pipeline to the
+core package.
+
+### 7. Contribute to training examples
+
+Diffusers examples are a collection of training scripts that reside in [examples](https://github.com/huggingface/diffusers/tree/main/examples).
+
+We support two types of training examples:
+
+- Official training examples
+- Research training examples
+
+Research training examples are located in [examples/research_projects](https://github.com/huggingface/diffusers/tree/main/examples/research_projects) whereas official training examples include all folders under [examples](https://github.com/huggingface/diffusers/tree/main/examples) except the `research_projects` and `community` folders.
+The official training examples are maintained by the Diffusers' core maintainers whereas the research training examples are maintained by the community.
+This is because of the same reasons put forward in [6. Contribute a community pipeline](#6-contribute-a-community-pipeline) for official pipelines vs. community pipelines: It is not feasible for the core maintainers to maintain all possible training methods for diffusion models.
+If the Diffusers core maintainers and the community consider a certain training paradigm to be too experimental or not popular enough, the corresponding training code should be put in the `research_projects` folder and maintained by the author.
+
+Both official training and research examples consist of a directory that contains one or more training scripts, a `requirements.txt` file, and a `README.md` file. In order for the user to make use of the
+training examples, it is required to clone the repository:
+
+```bash
+git clone https://github.com/huggingface/diffusers
+```
+
+as well as to install all additional dependencies required for training:
+
+```bash
+cd diffusers
+pip install -r examples/<your-example-folder>/requirements.txt
+```
+
+Therefore when adding an example, the `requirements.txt` file shall define all pip dependencies required for your training example so that once all those are installed, the user can run the example's training script. See, for example, the [DreamBooth `requirements.txt` file](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/requirements.txt).
+
+Training examples of the Diffusers library should adhere to the following philosophy:
+- All the code necessary to run the examples should be found in a single Python file.
+- One should be able to run the example from the command line with `python <your-example>.py --args`.
+- Examples should be kept simple and serve as **an example** on how to use Diffusers for training. The purpose of example scripts is **not** to create state-of-the-art diffusion models, but rather to reproduce known training schemes without adding too much custom logic. As a byproduct of this point, our examples also strive to serve as good educational materials.
+
+To contribute an example, it is highly recommended to look at already existing examples such as [dreambooth](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth.py) to get an idea of how they should look like.
+We strongly advise contributors to make use of the [Accelerate library](https://github.com/huggingface/accelerate) as it's tightly integrated
+with Diffusers.
+Once an example script works, please make sure to add a comprehensive `README.md` that states how to use the example exactly. This README should include:
+- An example command on how to run the example script as shown [here e.g.](https://github.com/huggingface/diffusers/tree/main/examples/dreambooth#running-locally-with-pytorch).
+- A link to some training results (logs, models, ...) that show what the user can expect as shown [here e.g.](https://api.wandb.ai/report/patrickvonplaten/xm6cd5q5).
+- If you are adding a non-official/research training example, **please don't forget** to add a sentence that you are maintaining this training example which includes your git handle as shown [here](https://github.com/huggingface/diffusers/tree/main/examples/research_projects/intel_opts#diffusers-examples-with-intel-optimizations).
+
+If you are contributing to the official training examples, please also make sure to add a test to [examples/test_examples.py](https://github.com/huggingface/diffusers/blob/main/examples/test_examples.py). This is not necessary for non-official training examples.
+
+### 8. Fixing a "Good second issue"
+
+*Good second issues* are marked by the [Good second issue](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22Good+second+issue%22) label. Good second issues are
+usually more complicated to solve than [Good first issues](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22good+first+issue%22).
+The issue description usually gives less guidance on how to fix the issue and requires
+a decent understanding of the library by the interested contributor.
+If you are interested in tackling a good second issue, feel free to open a PR to fix it and link the PR to the issue. If you see that a PR has already been opened for this issue but did not get merged, have a look to understand why it wasn't merged and try to open an improved PR.
+Good second issues are usually more difficult to get merged compared to good first issues, so don't hesitate to ask for help from the core maintainers. If your PR is almost finished the core maintainers can also jump into your PR and commit to it in order to get it merged.
+
+### 9. Adding pipelines, models, schedulers
+
+Pipelines, models, and schedulers are the most important pieces of the Diffusers library.
+They provide easy access to state-of-the-art diffusion technologies and thus allow the community to
+build powerful generative AI applications.
+
+By adding a new model, pipeline, or scheduler you might enable a new powerful use case for any of the user interfaces relying on Diffusers which can be of immense value for the whole generative AI ecosystem.
+
+Diffusers has a couple of open feature requests for all three components - feel free to gloss over them
+if you don't know yet what specific component you would like to add:
+- [Model or pipeline](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22New+pipeline%2Fmodel%22)
+- [Scheduler](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22New+scheduler%22)
+
+Before adding any of the three components, it is strongly recommended that you give the [Philosophy guide](https://github.com/huggingface/diffusers/blob/main/PHILOSOPHY.md) a read to better understand the design of any of the three components. Please be aware that
+we cannot merge model, scheduler, or pipeline additions that strongly diverge from our design philosophy
+as it will lead to API inconsistencies. If you fundamentally disagree with a design choice, please
+open a [Feedback issue](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=&template=feedback.md&title=) instead so that it can be discussed whether a certain design
+pattern/design choice shall be changed everywhere in the library and whether we shall update our design philosophy. Consistency across the library is very important for us.
+
+Please make sure to add links to the original codebase/paper to the PR and ideally also ping the
+original author directly on the PR so that they can follow the progress and potentially help with questions.
+
+If you are unsure or stuck in the PR, don't hesitate to leave a message to ask for a first review or help.
+
+## How to write a good issue
+
+**The better your issue is written, the higher the chances that it will be quickly resolved.**
+
+1. Make sure that you've used the correct template for your issue. You can pick between *Bug Report*, *Feature Request*, *Feedback about API Design*, *New model/pipeline/scheduler addition*, *Forum*, or a blank issue. Make sure to pick the correct one when opening [a new issue](https://github.com/huggingface/diffusers/issues/new/choose).
+2. **Be precise**: Give your issue a fitting title. Try to formulate your issue description as simple as possible. The more precise you are when submitting an issue, the less time it takes to understand the issue and potentially solve it. Make sure to open an issue for one issue only and not for multiple issues. If you found multiple issues, simply open multiple issues. If your issue is a bug, try to be as precise as possible about what bug it is - you should not just write "Error in diffusers".
+3. **Reproducibility**: No reproducible code snippet == no solution. If you encounter a bug, maintainers **have to be able to reproduce** it. Make sure that you include a code snippet that can be copy-pasted into a Python interpreter to reproduce the issue. Make sure that your code snippet works, *i.e.* that there are no missing imports or missing links to images, ... Your issue should contain an error message **and** a code snippet that can be copy-pasted without any changes to reproduce the exact same error message. If your issue is using local model weights or local data that cannot be accessed by the reader, the issue cannot be solved. If you cannot share your data or model, try to make a dummy model or dummy data.
+4. **Minimalistic**: Try to help the reader as much as you can to understand the issue as quickly as possible by staying as concise as possible. Remove all code / all information that is irrelevant to the issue. If you have found a bug, try to create the easiest code example you can to demonstrate your issue, do not just dump your whole workflow into the issue as soon as you have found a bug. E.g., if you train a model and get an error at some point during the training, you should first try to understand what part of the training code is responsible for the error and try to reproduce it with a couple of lines. Try to use dummy data instead of full datasets.
+5. Add links. If you are referring to a certain naming, method, or model make sure to provide a link so that the reader can better understand what you mean. If you are referring to a specific PR or issue, make sure to link it to your issue. Do not assume that the reader knows what you are talking about. The more links you add to your issue the better.
+6. Formatting. Make sure to nicely format your issue by formatting code into Python code syntax, and error messages into normal code syntax. See the [official GitHub formatting docs](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax) for more information.
+7. Think of your issue not as a ticket to be solved, but rather as a beautiful entry to a well-written encyclopedia. Every added issue is a contribution to publicly available knowledge. By adding a nicely written issue you not only make it easier for maintainers to solve your issue, but you are helping the whole community to better understand a certain aspect of the library.
+
+## How to write a good PR
+
+1. Be a chameleon. Understand existing design patterns and syntax and make sure your code additions flow seamlessly into the existing code base. Pull requests that significantly diverge from existing design patterns or user interfaces will not be merged.
+2. Be laser focused. A pull request should solve one problem and one problem only. Make sure to not fall into the trap of "also fixing another problem while we're adding it". It is much more difficult to review pull requests that solve multiple, unrelated problems at once.
+3. If helpful, try to add a code snippet that displays an example of how your addition can be used.
+4. The title of your pull request should be a summary of its contribution.
+5. If your pull request addresses an issue, please mention the issue number in
+the pull request description to make sure they are linked (and people
+consulting the issue know you are working on it);
+6. To indicate a work in progress please prefix the title with `[WIP]`. These
+are useful to avoid duplicated work, and to differentiate it from PRs ready
+to be merged;
+7. Try to formulate and format your text as explained in [How to write a good issue](#how-to-write-a-good-issue).
+8. Make sure existing tests pass;
+9. Add high-coverage tests. No quality testing = no merge.
+- If you are adding new `@slow` tests, make sure they pass using
+`RUN_SLOW=1 python -m pytest tests/test_my_new_model.py`.
+CircleCI does not run the slow tests, but GitHub Actions does every night!
+10. All public methods must have informative docstrings that work nicely with markdown. See [`pipeline_latent_diffusion.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py) for an example.
+11. Due to the rapidly growing repository, it is important to make sure that no files that would significantly weigh down the repository are added. This includes images, videos, and other non-text files. We prefer to leverage a hf.co hosted `dataset` like
+[`hf-internal-testing`](https://huggingface.co/hf-internal-testing) or [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images) to place these files.
+If an external contribution, feel free to add the images to your PR and ask a Hugging Face member to migrate your images
+to this dataset.
+
+## How to open a PR
+
+Before writing code, we strongly advise you to search through the existing PRs or
+issues to make sure that nobody is already working on the same thing. If you are
+unsure, it is always a good idea to open an issue to get some feedback.
+
+You will need basic `git` proficiency to be able to contribute to
+🧨 Diffusers. `git` is not the easiest tool to use but it has the greatest
+manual. Type `git --help` in a shell and enjoy. If you prefer books, [Pro
+Git](https://git-scm.com/book/en/v2) is a very good reference.
+
+Follow these steps to start contributing ([supported Python versions](https://github.com/huggingface/diffusers/blob/42f25d601a910dceadaee6c44345896b4cfa9928/setup.py#L270)):
+
+1. Fork the [repository](https://github.com/huggingface/diffusers) by
+clicking on the 'Fork' button on the repository's page. This creates a copy of the code
+under your GitHub user account.
+
+2. Clone your fork to your local disk, and add the base repository as a remote:
+
+ ```bash
+ $ git clone git@github.com:<your GitHub handle>/diffusers.git
+ $ cd diffusers
+ $ git remote add upstream https://github.com/huggingface/diffusers.git
+ ```
+
+3. Create a new branch to hold your development changes:
+
+ ```bash
+ $ git checkout -b a-descriptive-name-for-my-changes
+ ```
+
+**Do not** work on the `main` branch.
+
+4. Set up a development environment by running the following command in a virtual environment:
+
+ ```bash
+ $ pip install -e ".[dev]"
+ ```
+
+If you have already cloned the repo, you might need to `git pull` to get the most recent changes in the
+library.
+
+5. Develop the features on your branch.
+
+As you work on the features, you should make sure that the test suite
+passes. You should run the tests impacted by your changes like this:
+
+ ```bash
+ $ pytest tests/<TEST_TO_RUN>.py
+ ```
+
+Before you run the tests, please make sure you install the dependencies required for testing. You can do so
+with this command:
+
+ ```bash
+ $ pip install -e ".[test]"
+ ```
+
+You can also run the full test suite with the following command, but it takes
+a beefy machine to produce a result in a decent amount of time now that
+Diffusers has grown a lot. Here is the command for it:
+
+ ```bash
+ $ make test
+ ```
+
+🧨 Diffusers relies on `ruff` and `isort` to format its source code
+consistently. After you make changes, apply automatic style corrections and code verifications
+that can't be automated in one go with:
+
+ ```bash
+ $ make style
+ ```
+
+🧨 Diffusers also uses `ruff` and a few custom scripts to check for coding mistakes. Quality
+control runs in CI, however, you can also run the same checks with:
+
+ ```bash
+ $ make quality
+ ```
+
+Once you're happy with your changes, add changed files using `git add` and
+make a commit with `git commit` to record your changes locally:
+
+ ```bash
+ $ git add modified_file.py
+ $ git commit -m "A descriptive message about your changes."
+ ```
+
+It is a good idea to sync your copy of the code with the original
+repository regularly. This way you can quickly account for changes:
+
+ ```bash
+ $ git pull upstream main
+ ```
+
+Push the changes to your account using:
+
+ ```bash
+ $ git push -u origin a-descriptive-name-for-my-changes
+ ```
+
+6. Once you are satisfied, go to the
+webpage of your fork on GitHub. Click on 'Pull request' to send your changes
+to the project maintainers for review.
+
+7. It's ok if maintainers ask you for changes. It happens to core contributors
+too! So everyone can see the changes in the Pull request, work in your local
+branch and push the changes to your fork. They will automatically appear in
+the pull request.
+
+### Tests
+
+An extensive test suite is included to test the library behavior and several examples. Library tests can be found in
+the [tests folder](https://github.com/huggingface/diffusers/tree/main/tests).
+
+We like `pytest` and `pytest-xdist` because it's faster. From the root of the
+repository, here's how to run tests with `pytest` for the library:
+
+```bash
+$ python -m pytest -n auto --dist=loadfile -s -v ./tests/
+```
+
+In fact, that's how `make test` is implemented!
+
+You can specify a smaller set of tests in order to test only the feature
+you're working on.
+
+By default, slow tests are skipped. Set the `RUN_SLOW` environment variable to
+`yes` to run them. This will download many gigabytes of models — make sure you
+have enough disk space and a good Internet connection, or a lot of patience!
+
+```bash
+$ RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./tests/
+```
+
+`unittest` is fully supported, here's how to run tests with it:
+
+```bash
+$ python -m unittest discover -s tests -t . -v
+$ python -m unittest discover -s examples -t examples -v
+```
+
+### Syncing forked main with upstream (HuggingFace) main
+
+To avoid pinging the upstream repository which adds reference notes to each upstream PR and sends unnecessary notifications to the developers involved in these PRs,
+when syncing the main branch of a forked repository, please, follow these steps:
+1. When possible, avoid syncing with the upstream using a branch and PR on the forked repository. Instead, merge directly into the forked main.
+2. If a PR is absolutely necessary, use the following steps after checking out your branch:
+```bash
+$ git checkout -b your-branch-for-syncing
+$ git pull --squash --no-commit upstream main
+$ git commit -m '<your message without GitHub references>'
+$ git push --set-upstream origin your-branch-for-syncing
+```
+
+### Style guide
+
+For documentation strings, 🧨 Diffusers follows the [Google style](https://google.github.io/styleguide/pyguide.html).

docs/source/en/_toctree.yml

@@ -496,8 +496,6 @@
         title: Bria 3.2
       - local: api/pipelines/bria_fibo
         title: Bria Fibo
-      - local: api/pipelines/bria_fibo_edit
-        title: Bria Fibo Edit
       - local: api/pipelines/chroma
         title: Chroma
       - local: api/pipelines/cogview3

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

@@ -1,33 +0,0 @@
-<!--Copyright 2025 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.
--->
-
-# Bria Fibo Edit
-
-Fibo Edit is an 8B parameter image-to-image model that introduces a new paradigm of structured control, operating on JSON inputs paired with source images to enable deterministic and repeatable editing workflows.
-Featuring native masking for granular precision, it moves beyond simple prompt-based diffusion to offer explicit, interpretable control optimized for production environments.
-Its lightweight architecture is designed for deep customization, empowering researchers to build specialized "Edit" models for domain-specific tasks while delivering top-tier aesthetic quality
-
-## Usage
-_As the model is gated, before using it with diffusers you first need to go to the [Bria Fibo Hugging Face page](https://huggingface.co/briaai/Fibo-Edit), fill in the form and accept the gate. Once you are in, you need to login so that your system knows you’ve accepted the gate._
-
-Use the command below to log in:
-
-```bash
-hf auth login
-```
-
-
-## BriaFiboEditPipeline
-
-[[autodoc]] BriaFiboEditPipeline
-	- all
-	- __call__

src/diffusers/__init__.py

@@ -457,7 +457,6 @@ else:
             "AuraFlowPipeline",
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
-            "BriaFiboEditPipeline",
             "BriaFiboPipeline",
             "BriaPipeline",
             "ChromaImg2ImgPipeline",
@@ -1186,7 +1185,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDM2UNet2DConditionModel,
             AudioLDMPipeline,
             AuraFlowPipeline,
-            BriaFiboEditPipeline,
             BriaFiboPipeline,
             BriaPipeline,
             ChromaImg2ImgPipeline,

src/diffusers/models/attention_dispatch.py

@@ -235,10 +235,6 @@ class _AttentionBackendRegistry:
     def get_active_backend(cls):
         return cls._active_backend, cls._backends[cls._active_backend]
 
-    @classmethod
-    def set_active_backend(cls, backend: str):
-        cls._active_backend = backend
-
     @classmethod
     def list_backends(cls):
         return list(cls._backends.keys())
@@ -298,12 +294,12 @@ def attention_backend(backend: Union[str, AttentionBackendName] = AttentionBacke
     _maybe_download_kernel_for_backend(backend)
 
     old_backend = _AttentionBackendRegistry._active_backend
-    _AttentionBackendRegistry.set_active_backend(backend)
+    _AttentionBackendRegistry._active_backend = backend
 
     try:
         yield
     finally:
-        _AttentionBackendRegistry.set_active_backend(old_backend)
+        _AttentionBackendRegistry._active_backend = old_backend
 
 
 def dispatch_attention_fn(
@@ -352,7 +348,6 @@ def dispatch_attention_fn(
             check(**kwargs)
 
     kwargs = {k: v for k, v in kwargs.items() if k in _AttentionBackendRegistry._supported_arg_names[backend_name]}
-
     return backend_fn(**kwargs)
 
 

src/diffusers/models/modeling_utils.py

@@ -599,7 +599,6 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         from .attention import AttentionModuleMixin
         from .attention_dispatch import (
             AttentionBackendName,
-            _AttentionBackendRegistry,
             _check_attention_backend_requirements,
             _maybe_download_kernel_for_backend,
         )
@@ -608,16 +607,6 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         from .attention_processor import Attention, MochiAttention
 
         logger.warning("Attention backends are an experimental feature and the API may be subject to change.")
-        attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
-
-        parallel_config_set = False
-        for module in self.modules():
-            if not isinstance(module, attention_classes):
-                continue
-            processor = module.processor
-            if getattr(processor, "_parallel_config", None) is not None:
-                parallel_config_set = True
-                break
 
         backend = backend.lower()
         available_backends = {x.value for x in AttentionBackendName.__members__.values()}
@@ -625,17 +614,10 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))
 
         backend = AttentionBackendName(backend)
-        if parallel_config_set and not _AttentionBackendRegistry._is_context_parallel_available(backend):
-            compatible_backends = sorted(_AttentionBackendRegistry._supports_context_parallel)
-            raise ValueError(
-                f"Context parallelism is enabled but current attention backend '{backend.value}' "
-                f"does not support context parallelism. "
-                f"Please set a compatible attention backend: {compatible_backends} using `model.set_attention_backend()`."
-            )
-
         _check_attention_backend_requirements(backend)
         _maybe_download_kernel_for_backend(backend)
 
+        attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
         for module in self.modules():
             if not isinstance(module, attention_classes):
                 continue
@@ -644,9 +626,6 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 continue
             processor._attention_backend = backend
 
-        # Important to set the active backend so that it propagates gracefully throughout.
-        _AttentionBackendRegistry.set_active_backend(backend)
-
     def reset_attention_backend(self) -> None:
         """
         Resets the attention backend for the model. Following calls to `forward` will use the environment default, if
@@ -1559,7 +1538,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                         f"Context parallelism is enabled but the attention processor '{processor.__class__.__name__}' "
                         f"is using backend '{attention_backend.value}' which does not support context parallelism. "
                         f"Please set a compatible attention backend: {compatible_backends} using `model.set_attention_backend()` before "
-                        f"calling `model.enable_parallelism()`."
+                        f"calling `enable_parallelism()`."
                     )
 
                 # All modules use the same attention processor and backend. We don't need to

src/diffusers/models/transformers/transformer_flux.py

@@ -22,7 +22,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import apply_lora_scale, logging
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
 from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
 from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
@@ -634,7 +634,6 @@ class FluxTransformer2DModel(
 
         self.gradient_checkpointing = False
 
-    @apply_lora_scale("joint_attention_kwargs")
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -676,6 +675,20 @@ class FluxTransformer2DModel(
             If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
             `tuple` where the first element is the sample tensor.
         """
+        if joint_attention_kwargs is not None:
+            joint_attention_kwargs = joint_attention_kwargs.copy()
+            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
 
         hidden_states = self.x_embedder(hidden_states)
 
@@ -772,6 +785,10 @@ class FluxTransformer2DModel(
         hidden_states = self.norm_out(hidden_states, temb)
         output = self.proj_out(hidden_states)
 
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
         if not return_dict:
             return (output,)
 

src/diffusers/pipelines/__init__.py

@@ -129,7 +129,7 @@ else:
         "AnimateDiffVideoToVideoControlNetPipeline",
     ]
     _import_structure["bria"] = ["BriaPipeline"]
-    _import_structure["bria_fibo"] = ["BriaFiboPipeline", "BriaFiboEditPipeline"]
+    _import_structure["bria_fibo"] = ["BriaFiboPipeline"]
     _import_structure["flux2"] = ["Flux2Pipeline", "Flux2KleinPipeline"]
     _import_structure["flux"] = [
         "FluxControlPipeline",
@@ -597,7 +597,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .aura_flow import AuraFlowPipeline
         from .blip_diffusion import BlipDiffusionPipeline
         from .bria import BriaPipeline
-        from .bria_fibo import BriaFiboEditPipeline, BriaFiboPipeline
+        from .bria_fibo import BriaFiboPipeline
         from .chroma import ChromaImg2ImgPipeline, ChromaInpaintPipeline, ChromaPipeline
         from .chronoedit import ChronoEditPipeline
         from .cogvideo import (

src/diffusers/pipelines/bria_fibo/__init__.py

@@ -23,8 +23,6 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_bria_fibo"] = ["BriaFiboPipeline"]
-    _import_structure["pipeline_bria_fibo_edit"] = ["BriaFiboEditPipeline"]
-
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -35,7 +33,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
         from .pipeline_bria_fibo import BriaFiboPipeline
-        from .pipeline_bria_fibo_edit import BriaFiboEditPipeline
 
 else:
     import sys

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py

@@ -84,6 +84,7 @@ EXAMPLE_DOC_STRING = """
         >>> from diffusers import ControlNetModel, StableDiffusionXLControlNetImg2ImgPipeline, AutoencoderKL
         >>> from diffusers.utils import load_image
 
+
         >>> depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda")
         >>> feature_extractor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas")
         >>> controlnet = ControlNetModel.from_pretrained(

src/diffusers/pipelines/hidream_image/pipeline_hidream_image.py

@@ -53,6 +53,7 @@ EXAMPLE_DOC_STRING = """
         >>> from transformers import AutoTokenizer, LlamaForCausalLM
         >>> from diffusers import HiDreamImagePipeline
 
+
         >>> tokenizer_4 = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")
         >>> text_encoder_4 = LlamaForCausalLM.from_pretrained(
         ...     "meta-llama/Meta-Llama-3.1-8B-Instruct",

src/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py

@@ -85,6 +85,7 @@ EXAMPLE_DOC_STRING = """
         >>> from diffusers import ControlNetModel, StableDiffusionXLControlNetPAGImg2ImgPipeline, AutoencoderKL
         >>> from diffusers.utils import load_image
 
+
         >>> depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda")
         >>> feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas")
         >>> controlnet = ControlNetModel.from_pretrained(

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py

@@ -459,6 +459,7 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
         >>> from diffusers import StableDiffusionLatentUpscalePipeline, StableDiffusionPipeline
         >>> import torch
 
+
         >>> pipeline = StableDiffusionPipeline.from_pretrained(
         ...     "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16
         ... )

src/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py

@@ -15,7 +15,7 @@
 # DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver and https://github.com/NVlabs/edm
 
 import math
-from typing import List, Literal, Optional, Tuple, Union
+from typing import List, Optional, Tuple, Union
 
 import numpy as np
 import torch
@@ -36,30 +36,27 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
     methods the library implements for all schedulers such as loading and saving.
 
     Args:
-        sigma_min (`float`, defaults to `0.3`):
+        sigma_min (`float`, *optional*, defaults to 0.3):
             Minimum noise magnitude in the sigma schedule. This was set to 0.3 in Stable Audio Open [1].
-        sigma_max (`float`, defaults to `500`):
+        sigma_max (`float`, *optional*, defaults to 500):
             Maximum noise magnitude in the sigma schedule. This was set to 500 in Stable Audio Open [1].
-        sigma_data (`float`, defaults to `1.0`):
+        sigma_data (`float`, *optional*, defaults to 1.0):
             The standard deviation of the data distribution. This is set to 1.0 in Stable Audio Open [1].
-        sigma_schedule (`str`, defaults to `"exponential"`):
-            Sigma schedule to compute the `sigmas`. Must be one of `"exponential"` or `"karras"`. The exponential
-            schedule was incorporated in [stabilityai/cosxl](https://huggingface.co/stabilityai/cosxl). The Karras
-            schedule is introduced in the [EDM](https://huggingface.co/papers/2206.00364) paper.
-        num_train_timesteps (`int`, defaults to `1000`):
+        sigma_schedule (`str`, *optional*, defaults to `exponential`):
+            Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper
+            (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential
+            schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl.
+        num_train_timesteps (`int`, defaults to 1000):
             The number of diffusion steps to train the model.
-        solver_order (`int`, defaults to `2`):
+        solver_order (`int`, defaults to 2):
             The DPMSolver order which can be `1` or `2`. It is recommended to use `solver_order=2`.
-        prediction_type (`str`, defaults to `"v_prediction"`):
-            Prediction type of the scheduler function. Must be one of `"epsilon"` (predicts the noise of the diffusion
-            process), `"sample"` (directly predicts the noisy sample), or `"v_prediction"` (see section 2.4 of [Imagen
+        prediction_type (`str`, defaults to `v_prediction`, *optional*):
+            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+            `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
             Video](https://huggingface.co/papers/2210.02303) paper).
-        rho (`float`, defaults to `7.0`):
-            The parameter for calculating the Karras sigma schedule from the EDM
-            [paper](https://huggingface.co/papers/2206.00364).
-        solver_type (`str`, defaults to `"midpoint"`):
-            Solver type for the second-order solver. Must be one of `"midpoint"` or `"heun"`. The solver type slightly
-            affects the sample quality, especially for a small number of steps. It is recommended to use `"midpoint"`.
+        solver_type (`str`, defaults to `midpoint`):
+            Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
+            sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
         lower_order_final (`bool`, defaults to `True`):
             Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
             stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
@@ -68,9 +65,8 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
             steps, but sometimes may result in blurring.
         final_sigmas_type (`str`, defaults to `"zero"`):
-            The final `sigma` value for the noise schedule during the sampling process. Must be one of `"zero"` or
-            `"sigma_min"`. If `"sigma_min"`, the final sigma is the same as the last sigma in the training schedule. If
-            `"zero"`, the final sigma is set to 0.
+            The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final
+            sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
     """
 
     _compatibles = []
@@ -82,16 +78,16 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         sigma_min: float = 0.3,
         sigma_max: float = 500,
         sigma_data: float = 1.0,
-        sigma_schedule: Literal["exponential", "karras"] = "exponential",
+        sigma_schedule: str = "exponential",
         num_train_timesteps: int = 1000,
         solver_order: int = 2,
-        prediction_type: Literal["epsilon", "sample", "v_prediction"] = "v_prediction",
+        prediction_type: str = "v_prediction",
         rho: float = 7.0,
-        solver_type: Literal["midpoint", "heun"] = "midpoint",
+        solver_type: str = "midpoint",
         lower_order_final: bool = True,
         euler_at_final: bool = False,
-        final_sigmas_type: Literal["zero", "sigma_min"] = "zero",
-    ) -> None:
+        final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
+    ):
         if solver_type not in ["midpoint", "heun"]:
             if solver_type in ["logrho", "bh1", "bh2"]:
                 self.register_to_config(solver_type="midpoint")
@@ -117,40 +113,26 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     @property
-    def init_noise_sigma(self) -> float:
-        """
-        The standard deviation of the initial noise distribution.
-
-        Returns:
-            `float`:
-                The initial noise sigma value computed as `sqrt(sigma_max^2 + 1)`.
-        """
+    def init_noise_sigma(self):
+        # standard deviation of the initial noise distribution
         return (self.config.sigma_max**2 + 1) ** 0.5
 
     @property
-    def step_index(self) -> Optional[int]:
+    def step_index(self):
         """
         The index counter for current timestep. It will increase 1 after each scheduler step.
-
-        Returns:
-            `int` or `None`:
-                The current step index, or `None` if not yet initialized.
         """
         return self._step_index
 
     @property
-    def begin_index(self) -> Optional[int]:
+    def begin_index(self):
         """
         The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
-
-        Returns:
-            `int` or `None`:
-                The begin index, or `None` if not yet set.
         """
         return self._begin_index
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
-    def set_begin_index(self, begin_index: int = 0) -> None:
+    def set_begin_index(self, begin_index: int = 0):
         """
         Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
 
@@ -179,18 +161,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         scaled_sample = sample * c_in
         return scaled_sample
 
-    def precondition_noise(self, sigma: Union[float, torch.Tensor]) -> torch.Tensor:
-        """
-        Precondition the noise level by computing a normalized timestep representation.
-
-        Args:
-            sigma (`float` or `torch.Tensor`):
-                The sigma (noise level) value to precondition.
-
-        Returns:
-            `torch.Tensor`:
-                The preconditioned noise value computed as `atan(sigma) / pi * 2`.
-        """
+    def precondition_noise(self, sigma):
         if not isinstance(sigma, torch.Tensor):
             sigma = torch.tensor([sigma])
 
@@ -257,14 +228,12 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         self.is_scale_input_called = True
         return sample
 
-    def set_timesteps(
-        self, num_inference_steps: Optional[int] = None, device: Optional[Union[str, torch.device]] = None
-    ) -> None:
+    def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None):
         """
         Sets the discrete timesteps used for the diffusion chain (to be run before inference).
 
         Args:
-            num_inference_steps (`int`, *optional*):
+            num_inference_steps (`int`):
                 The number of diffusion steps used when generating samples with a pre-trained model.
             device (`str` or `torch.device`, *optional*):
                 The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
@@ -365,7 +334,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         return sigmas
 
     # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
-    def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray:
+    def _sigma_to_t(self, sigma, log_sigmas):
         """
         Convert sigma values to corresponding timestep values through interpolation.
 
@@ -401,19 +370,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         t = t.reshape(sigma.shape)
         return t
 
-    def _sigma_to_alpha_sigma_t(self, sigma: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        """
-        Convert sigma to alpha and sigma_t values for the diffusion process.
-
-        Args:
-            sigma (`torch.Tensor`):
-                The sigma (noise level) value.
-
-        Returns:
-            `Tuple[torch.Tensor, torch.Tensor]`:
-                A tuple containing `alpha_t` (always 1 since inputs are pre-scaled) and `sigma_t` (same as input
-                sigma).
-        """
+    def _sigma_to_alpha_sigma_t(self, sigma):
         alpha_t = torch.tensor(1)  # Inputs are pre-scaled before going into unet, so alpha_t = 1
         sigma_t = sigma
 
@@ -579,7 +536,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         return step_index
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index
-    def _init_step_index(self, timestep: Union[int, torch.Tensor]) -> None:
+    def _init_step_index(self, timestep):
         """
         Initialize the step_index counter for the scheduler.
 
@@ -600,7 +557,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         model_output: torch.Tensor,
         timestep: Union[int, torch.Tensor],
         sample: torch.Tensor,
-        generator: Optional[torch.Generator] = None,
+        generator=None,
         return_dict: bool = True,
     ) -> Union[SchedulerOutput, Tuple]:
         """
@@ -610,19 +567,20 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         Args:
             model_output (`torch.Tensor`):
                 The direct output from learned diffusion model.
-            timestep (`int` or `torch.Tensor`):
+            timestep (`int`):
                 The current discrete timestep in the diffusion chain.
             sample (`torch.Tensor`):
                 A current instance of a sample created by the diffusion process.
             generator (`torch.Generator`, *optional*):
                 A random number generator.
-            return_dict (`bool`, defaults to `True`):
+            return_dict (`bool`):
                 Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
 
         Returns:
             [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
                 If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
                 tuple is returned where the first element is the sample tensor.
+
         """
         if self.num_inference_steps is None:
             raise ValueError(
@@ -744,12 +702,5 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
         return c_in
 
-    def __len__(self) -> int:
-        """
-        Returns the number of training timesteps.
-
-        Returns:
-            `int`:
-                The number of training timesteps configured for the scheduler.
-        """
+    def __len__(self):
         return self.config.num_train_timesteps

src/diffusers/utils/__init__.py

@@ -130,7 +130,6 @@ from .loading_utils import get_module_from_name, get_submodule_by_name, load_ima
 from .logging import get_logger
 from .outputs import BaseOutput
 from .peft_utils import (
-    apply_lora_scale,
     check_peft_version,
     delete_adapter_layers,
     get_adapter_name,

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -587,21 +587,6 @@ class AuraFlowPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
-class BriaFiboEditPipeline(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 BriaFiboPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

src/diffusers/utils/peft_utils.py

@@ -16,7 +16,6 @@ PEFT utilities: Utilities related to peft library
 """
 
 import collections
-import functools
 import importlib
 from typing import Optional
 
@@ -276,59 +275,6 @@ def set_weights_and_activate_adapters(model, adapter_names, weights):
                 module.set_scale(adapter_name, get_module_weight(weight, module_name))
 
 
-def apply_lora_scale(kwargs_name: str = "joint_attention_kwargs"):
-    """
-    Decorator to automatically handle LoRA layer scaling/unscaling in forward methods.
-
-    This decorator extracts the `lora_scale` from the specified kwargs parameter, applies scaling before the forward
-    pass, and ensures unscaling happens after, even if an exception occurs.
-
-    Args:
-        kwargs_name (`str`, defaults to `"joint_attention_kwargs"`):
-            The name of the keyword argument that contains the LoRA scale. Common values include
-            "joint_attention_kwargs", "attention_kwargs", "cross_attention_kwargs", etc.
-    """
-
-    def decorator(forward_fn):
-        @functools.wraps(forward_fn)
-        def wrapper(self, *args, **kwargs):
-            from . import USE_PEFT_BACKEND
-
-            lora_scale = 1.0
-            attention_kwargs = kwargs.get(kwargs_name)
-
-            if attention_kwargs is not None:
-                attention_kwargs = attention_kwargs.copy()
-                kwargs[kwargs_name] = attention_kwargs
-                lora_scale = attention_kwargs.pop("scale", 1.0)
-            else:
-                if (
-                    not USE_PEFT_BACKEND
-                    and attention_kwargs is not None
-                    and attention_kwargs.get("scale", None) is not None
-                ):
-                    logger.warning(
-                        f"Passing `scale` via `{kwargs_name}` when not using the PEFT backend is ineffective."
-                    )
-
-            # Apply LoRA scaling if using PEFT backend
-            if USE_PEFT_BACKEND:
-                scale_lora_layers(self, lora_scale)
-
-            try:
-                # Execute the forward pass
-                result = forward_fn(self, *args, **kwargs)
-                return result
-            finally:
-                # Always unscale, even if forward pass raises an exception
-                if USE_PEFT_BACKEND:
-                    unscale_lora_layers(self, lora_scale)
-
-        return wrapper
-
-    return decorator
-
-
 def check_peft_version(min_version: str) -> None:
     r"""
     Checks if the version of PEFT is compatible.

tests/conftest.py

@@ -32,6 +32,22 @@ warnings.simplefilter(action="ignore", category=FutureWarning)
 
 def pytest_configure(config):
     config.addinivalue_line("markers", "big_accelerator: marks tests as requiring big accelerator resources")
+    config.addinivalue_line("markers", "lora: marks tests for LoRA/PEFT functionality")
+    config.addinivalue_line("markers", "ip_adapter: marks tests for IP Adapter functionality")
+    config.addinivalue_line("markers", "training: marks tests for training functionality")
+    config.addinivalue_line("markers", "attention: marks tests for attention processor functionality")
+    config.addinivalue_line("markers", "memory: marks tests for memory optimization functionality")
+    config.addinivalue_line("markers", "cpu_offload: marks tests for CPU offloading functionality")
+    config.addinivalue_line("markers", "group_offload: marks tests for group offloading functionality")
+    config.addinivalue_line("markers", "compile: marks tests for torch.compile functionality")
+    config.addinivalue_line("markers", "single_file: marks tests for single file checkpoint loading")
+    config.addinivalue_line("markers", "quantization: marks tests for quantization functionality")
+    config.addinivalue_line("markers", "bitsandbytes: marks tests for BitsAndBytes quantization functionality")
+    config.addinivalue_line("markers", "quanto: marks tests for Quanto quantization functionality")
+    config.addinivalue_line("markers", "torchao: marks tests for TorchAO quantization functionality")
+    config.addinivalue_line("markers", "gguf: marks tests for GGUF quantization functionality")
+    config.addinivalue_line("markers", "modelopt: marks tests for NVIDIA ModelOpt quantization functionality")
+    config.addinivalue_line("markers", "context_parallel: marks tests for context parallel inference functionality")
     config.addinivalue_line("markers", "slow: mark test as slow")
     config.addinivalue_line("markers", "nightly: mark test as nightly")
 

tests/models/testing_utils/__init__.py

@@ -0,0 +1,79 @@
+from .attention import AttentionTesterMixin
+from .cache import (
+    CacheTesterMixin,
+    FasterCacheConfigMixin,
+    FasterCacheTesterMixin,
+    FirstBlockCacheConfigMixin,
+    FirstBlockCacheTesterMixin,
+    PyramidAttentionBroadcastConfigMixin,
+    PyramidAttentionBroadcastTesterMixin,
+)
+from .common import BaseModelTesterConfig, ModelTesterMixin
+from .compile import TorchCompileTesterMixin
+from .ip_adapter import IPAdapterTesterMixin
+from .lora import LoraHotSwappingForModelTesterMixin, LoraTesterMixin
+from .memory import CPUOffloadTesterMixin, GroupOffloadTesterMixin, LayerwiseCastingTesterMixin, MemoryTesterMixin
+from .parallelism import ContextParallelTesterMixin
+from .quantization import (
+    BitsAndBytesCompileTesterMixin,
+    BitsAndBytesConfigMixin,
+    BitsAndBytesTesterMixin,
+    GGUFCompileTesterMixin,
+    GGUFConfigMixin,
+    GGUFTesterMixin,
+    ModelOptCompileTesterMixin,
+    ModelOptConfigMixin,
+    ModelOptTesterMixin,
+    QuantizationCompileTesterMixin,
+    QuantizationTesterMixin,
+    QuantoCompileTesterMixin,
+    QuantoConfigMixin,
+    QuantoTesterMixin,
+    TorchAoCompileTesterMixin,
+    TorchAoConfigMixin,
+    TorchAoTesterMixin,
+)
+from .single_file import SingleFileTesterMixin
+from .training import TrainingTesterMixin
+
+
+__all__ = [
+    "AttentionTesterMixin",
+    "BaseModelTesterConfig",
+    "BitsAndBytesCompileTesterMixin",
+    "BitsAndBytesConfigMixin",
+    "BitsAndBytesTesterMixin",
+    "CacheTesterMixin",
+    "ContextParallelTesterMixin",
+    "CPUOffloadTesterMixin",
+    "FasterCacheConfigMixin",
+    "FasterCacheTesterMixin",
+    "FirstBlockCacheConfigMixin",
+    "FirstBlockCacheTesterMixin",
+    "GGUFCompileTesterMixin",
+    "GGUFConfigMixin",
+    "GGUFTesterMixin",
+    "GroupOffloadTesterMixin",
+    "IPAdapterTesterMixin",
+    "LayerwiseCastingTesterMixin",
+    "LoraHotSwappingForModelTesterMixin",
+    "LoraTesterMixin",
+    "MemoryTesterMixin",
+    "ModelOptCompileTesterMixin",
+    "ModelOptConfigMixin",
+    "ModelOptTesterMixin",
+    "ModelTesterMixin",
+    "PyramidAttentionBroadcastConfigMixin",
+    "PyramidAttentionBroadcastTesterMixin",
+    "QuantizationCompileTesterMixin",
+    "QuantizationTesterMixin",
+    "QuantoCompileTesterMixin",
+    "QuantoConfigMixin",
+    "QuantoTesterMixin",
+    "SingleFileTesterMixin",
+    "TorchAoCompileTesterMixin",
+    "TorchAoConfigMixin",
+    "TorchAoTesterMixin",
+    "TorchCompileTesterMixin",
+    "TrainingTesterMixin",
+]

tests/models/testing_utils/attention.py

@@ -0,0 +1,181 @@
+# coding=utf-8
+# Copyright 2025 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 pytest
+import torch
+
+from diffusers.models.attention import AttentionModuleMixin
+from diffusers.models.attention_processor import (
+    AttnProcessor,
+)
+
+from ...testing_utils import (
+    assert_tensors_close,
+    backend_empty_cache,
+    is_attention,
+    torch_device,
+)
+
+
+@is_attention
+class AttentionTesterMixin:
+    """
+    Mixin class for testing attention processor and module functionality on models.
+
+    Tests functionality from AttentionModuleMixin including:
+        - Attention processor management (set/get)
+        - QKV projection fusion/unfusion
+        - Attention backends (XFormers, NPU, etc.)
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: attention
+        Use `pytest -m "not attention"` to skip these tests
+    """
+
+    def setup_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def teardown_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    @torch.no_grad()
+    def test_fuse_unfuse_qkv_projections(self, atol=1e-3, rtol=0):
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.eval()
+
+        if not hasattr(model, "fuse_qkv_projections"):
+            pytest.skip("Model does not support QKV projection fusion.")
+
+        output_before_fusion = model(**inputs_dict, return_dict=False)[0]
+
+        model.fuse_qkv_projections()
+
+        has_fused_projections = False
+        for module in model.modules():
+            if isinstance(module, AttentionModuleMixin):
+                if hasattr(module, "to_qkv") or hasattr(module, "to_kv"):
+                    has_fused_projections = True
+                    assert module.fused_projections, "fused_projections flag should be True"
+                    break
+
+        if has_fused_projections:
+            output_after_fusion = model(**inputs_dict, return_dict=False)[0]
+
+            assert_tensors_close(
+                output_before_fusion,
+                output_after_fusion,
+                atol=atol,
+                rtol=rtol,
+                msg="Output should not change after fusing projections",
+            )
+
+            model.unfuse_qkv_projections()
+
+            for module in model.modules():
+                if isinstance(module, AttentionModuleMixin):
+                    assert not hasattr(module, "to_qkv"), "to_qkv should be removed after unfusing"
+                    assert not hasattr(module, "to_kv"), "to_kv should be removed after unfusing"
+                    assert not module.fused_projections, "fused_projections flag should be False"
+
+            output_after_unfusion = model(**inputs_dict, return_dict=False)[0]
+
+            assert_tensors_close(
+                output_before_fusion,
+                output_after_unfusion,
+                atol=atol,
+                rtol=rtol,
+                msg="Output should match original after unfusing projections",
+            )
+
+    def test_get_set_processor(self):
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+
+        # Check if model has attention processors
+        if not hasattr(model, "attn_processors"):
+            pytest.skip("Model does not have attention processors.")
+
+        # Test getting processors
+        processors = model.attn_processors
+        assert isinstance(processors, dict), "attn_processors should return a dict"
+        assert len(processors) > 0, "Model should have at least one attention processor"
+
+        # Test that all processors can be retrieved via get_processor
+        for module in model.modules():
+            if isinstance(module, AttentionModuleMixin):
+                processor = module.get_processor()
+                assert processor is not None, "get_processor should return a processor"
+
+                # Test setting a new processor
+                new_processor = AttnProcessor()
+                module.set_processor(new_processor)
+                retrieved_processor = module.get_processor()
+                assert retrieved_processor is new_processor, "Retrieved processor should be the same as the one set"
+
+    def test_attention_processor_dict(self):
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+
+        if not hasattr(model, "set_attn_processor"):
+            pytest.skip("Model does not support setting attention processors.")
+
+        # Get current processors
+        current_processors = model.attn_processors
+
+        # Create a dict of new processors
+        new_processors = {key: AttnProcessor() for key in current_processors.keys()}
+
+        # Set processors using dict
+        model.set_attn_processor(new_processors)
+
+        # Verify all processors were set
+        updated_processors = model.attn_processors
+        for key in current_processors.keys():
+            assert type(updated_processors[key]) == AttnProcessor, f"Processor {key} should be AttnProcessor"
+
+    def test_attention_processor_count_mismatch_raises_error(self):
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+
+        if not hasattr(model, "set_attn_processor"):
+            pytest.skip("Model does not support setting attention processors.")
+
+        # Get current processors
+        current_processors = model.attn_processors
+
+        # Create a dict with wrong number of processors
+        wrong_processors = {list(current_processors.keys())[0]: AttnProcessor()}
+
+        # Verify error is raised
+        with pytest.raises(ValueError) as exc_info:
+            model.set_attn_processor(wrong_processors)
+
+        assert "number of processors" in str(exc_info.value).lower(), "Error should mention processor count mismatch"

tests/models/testing_utils/cache.py

@@ -0,0 +1,556 @@
+# coding=utf-8
+# Copyright 2025 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 pytest
+import torch
+
+from diffusers.hooks import FasterCacheConfig, FirstBlockCacheConfig, PyramidAttentionBroadcastConfig
+from diffusers.hooks.faster_cache import _FASTER_CACHE_BLOCK_HOOK, _FASTER_CACHE_DENOISER_HOOK
+from diffusers.hooks.first_block_cache import _FBC_BLOCK_HOOK, _FBC_LEADER_BLOCK_HOOK
+from diffusers.hooks.pyramid_attention_broadcast import _PYRAMID_ATTENTION_BROADCAST_HOOK
+from diffusers.models.cache_utils import CacheMixin
+
+from ...testing_utils import assert_tensors_close, backend_empty_cache, is_cache, torch_device
+
+
+def require_cache_mixin(func):
+    """Decorator to skip tests if model doesn't use CacheMixin."""
+
+    def wrapper(self, *args, **kwargs):
+        if not issubclass(self.model_class, CacheMixin):
+            pytest.skip(f"{self.model_class.__name__} does not use CacheMixin.")
+        return func(self, *args, **kwargs)
+
+    return wrapper
+
+
+class CacheTesterMixin:
+    """
+    Base mixin class providing common test implementations for cache testing.
+
+    Cache-specific mixins should:
+    1. Inherit from their respective config mixin (e.g., PyramidAttentionBroadcastConfigMixin)
+    2. Inherit from this mixin
+    3. Define the cache config to use for tests
+
+    Expected class attributes:
+        - model_class: The model class to test (must use CacheMixin)
+
+    Expected methods in test classes:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Optional overrides:
+        - cache_input_key: Property returning the input tensor key to vary between passes (default: "hidden_states")
+    """
+
+    @property
+    def cache_input_key(self):
+        return "hidden_states"
+
+    def setup_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def teardown_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def _get_cache_config(self):
+        """
+        Get the cache config for testing.
+        Should be implemented by subclasses.
+        """
+        raise NotImplementedError("Subclass must implement _get_cache_config")
+
+    def _get_hook_names(self):
+        """
+        Get the hook names to check for this cache type.
+        Should be implemented by subclasses.
+        Returns a list of hook name strings.
+        """
+        raise NotImplementedError("Subclass must implement _get_hook_names")
+
+    def _test_cache_enable_disable_state(self):
+        """Test that cache enable/disable updates the is_cache_enabled state correctly."""
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        # Initially cache should not be enabled
+        assert not model.is_cache_enabled, "Cache should not be enabled initially."
+
+        config = self._get_cache_config()
+
+        # Enable cache
+        model.enable_cache(config)
+        assert model.is_cache_enabled, "Cache should be enabled after enable_cache()."
+
+        # Disable cache
+        model.disable_cache()
+        assert not model.is_cache_enabled, "Cache should not be enabled after disable_cache()."
+
+    def _test_cache_double_enable_raises_error(self):
+        """Test that enabling cache twice raises an error."""
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        config = self._get_cache_config()
+
+        model.enable_cache(config)
+
+        # Trying to enable again should raise ValueError
+        with pytest.raises(ValueError, match="Caching has already been enabled"):
+            model.enable_cache(config)
+
+        # Cleanup
+        model.disable_cache()
+
+    def _test_cache_hooks_registered(self):
+        """Test that cache hooks are properly registered and removed."""
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        config = self._get_cache_config()
+        hook_names = self._get_hook_names()
+
+        model.enable_cache(config)
+
+        # Check that at least one hook was registered
+        hook_count = 0
+        for module in model.modules():
+            if hasattr(module, "_diffusers_hook"):
+                for hook_name in hook_names:
+                    hook = module._diffusers_hook.get_hook(hook_name)
+                    if hook is not None:
+                        hook_count += 1
+
+        assert hook_count > 0, f"At least one cache hook should be registered. Hook names: {hook_names}"
+
+        # Disable and verify hooks are removed
+        model.disable_cache()
+
+        hook_count_after = 0
+        for module in model.modules():
+            if hasattr(module, "_diffusers_hook"):
+                for hook_name in hook_names:
+                    hook = module._diffusers_hook.get_hook(hook_name)
+                    if hook is not None:
+                        hook_count_after += 1
+
+        assert hook_count_after == 0, "Cache hooks should be removed after disable_cache()."
+
+    @torch.no_grad()
+    def _test_cache_inference(self):
+        """Test that model can run inference with cache enabled."""
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+
+        config = self._get_cache_config()
+
+        model.enable_cache(config)
+
+        # First pass populates the cache
+        _ = model(**inputs_dict, return_dict=False)[0]
+
+        # Create modified inputs for second pass (vary input tensor to simulate denoising)
+        inputs_dict_step2 = inputs_dict.copy()
+        if self.cache_input_key in inputs_dict_step2:
+            inputs_dict_step2[self.cache_input_key] = inputs_dict_step2[self.cache_input_key] + torch.randn_like(
+                inputs_dict_step2[self.cache_input_key]
+            )
+
+        # Second pass uses cached attention with different inputs (produces approximated output)
+        output_with_cache = model(**inputs_dict_step2, return_dict=False)[0]
+
+        assert output_with_cache is not None, "Model output should not be None with cache enabled."
+        assert not torch.isnan(output_with_cache).any(), "Model output contains NaN with cache enabled."
+
+        # Run same inputs without cache to compare
+        model.disable_cache()
+        output_without_cache = model(**inputs_dict_step2, return_dict=False)[0]
+
+        # Cached output should be different from non-cached output (due to approximation)
+        assert not torch.allclose(output_without_cache, output_with_cache, atol=1e-5), (
+            "Cached output should be different from non-cached output due to cache approximation."
+        )
+
+    @torch.no_grad()
+    def _test_cache_context_manager(self, atol=1e-5, rtol=0):
+        """Test the cache_context context manager properly isolates cache state."""
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+
+        config = self._get_cache_config()
+        model.enable_cache(config)
+
+        # Run inference in first context
+        with model.cache_context("context_1"):
+            output_ctx1 = model(**inputs_dict, return_dict=False)[0]
+
+        # Run same inference in second context (cache should be reset)
+        with model.cache_context("context_2"):
+            output_ctx2 = model(**inputs_dict, return_dict=False)[0]
+
+        # Both contexts should produce the same output (first pass in each)
+        assert_tensors_close(
+            output_ctx1,
+            output_ctx2,
+            atol=atol,
+            rtol=rtol,
+            msg="First pass in different cache contexts should produce the same output.",
+        )
+
+        model.disable_cache()
+
+    @torch.no_grad()
+    def _test_reset_stateful_cache(self):
+        """Test that _reset_stateful_cache resets the cache state."""
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+
+        config = self._get_cache_config()
+
+        model.enable_cache(config)
+
+        _ = model(**inputs_dict, return_dict=False)[0]
+
+        model._reset_stateful_cache()
+
+        model.disable_cache()
+
+
+@is_cache
+class PyramidAttentionBroadcastConfigMixin:
+    """
+    Base mixin providing PyramidAttentionBroadcast cache config.
+
+    Expected class attributes:
+        - model_class: The model class to test (must use CacheMixin)
+    """
+
+    # Default PAB config - can be overridden by subclasses
+    PAB_CONFIG = {
+        "spatial_attention_block_skip_range": 2,
+    }
+
+    # Store timestep for callback (must be within default range (100, 800) for skipping to trigger)
+    _current_timestep = 500
+
+    def _get_cache_config(self):
+        config_kwargs = self.PAB_CONFIG.copy()
+        config_kwargs["current_timestep_callback"] = lambda: self._current_timestep
+        return PyramidAttentionBroadcastConfig(**config_kwargs)
+
+    def _get_hook_names(self):
+        return [_PYRAMID_ATTENTION_BROADCAST_HOOK]
+
+
+@is_cache
+class PyramidAttentionBroadcastTesterMixin(PyramidAttentionBroadcastConfigMixin, CacheTesterMixin):
+    """
+    Mixin class for testing PyramidAttentionBroadcast caching on models.
+
+    Expected class attributes:
+        - model_class: The model class to test (must use CacheMixin)
+
+    Expected methods to be implemented by subclasses:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: cache
+        Use `pytest -m "not cache"` to skip these tests
+    """
+
+    @require_cache_mixin
+    def test_pab_cache_enable_disable_state(self):
+        self._test_cache_enable_disable_state()
+
+    @require_cache_mixin
+    def test_pab_cache_double_enable_raises_error(self):
+        self._test_cache_double_enable_raises_error()
+
+    @require_cache_mixin
+    def test_pab_cache_hooks_registered(self):
+        self._test_cache_hooks_registered()
+
+    @require_cache_mixin
+    def test_pab_cache_inference(self):
+        self._test_cache_inference()
+
+    @require_cache_mixin
+    def test_pab_cache_context_manager(self):
+        self._test_cache_context_manager()
+
+    @require_cache_mixin
+    def test_pab_reset_stateful_cache(self):
+        self._test_reset_stateful_cache()
+
+
+@is_cache
+class FirstBlockCacheConfigMixin:
+    """
+    Base mixin providing FirstBlockCache config.
+
+    Expected class attributes:
+        - model_class: The model class to test (must use CacheMixin)
+    """
+
+    # Default FBC config - can be overridden by subclasses
+    # Higher threshold makes FBC more aggressive about caching (skips more often)
+    FBC_CONFIG = {
+        "threshold": 1.0,
+    }
+
+    def _get_cache_config(self):
+        return FirstBlockCacheConfig(**self.FBC_CONFIG)
+
+    def _get_hook_names(self):
+        return [_FBC_LEADER_BLOCK_HOOK, _FBC_BLOCK_HOOK]
+
+
+@is_cache
+class FirstBlockCacheTesterMixin(FirstBlockCacheConfigMixin, CacheTesterMixin):
+    """
+    Mixin class for testing FirstBlockCache on models.
+
+    Expected class attributes:
+        - model_class: The model class to test (must use CacheMixin)
+
+    Expected methods to be implemented by subclasses:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: cache
+        Use `pytest -m "not cache"` to skip these tests
+    """
+
+    @torch.no_grad()
+    def _test_cache_inference(self):
+        """Test that model can run inference with FBC cache enabled (requires cache_context)."""
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+
+        config = self._get_cache_config()
+        model.enable_cache(config)
+
+        # FBC requires cache_context to be set for inference
+        with model.cache_context("fbc_test"):
+            # First pass populates the cache
+            _ = model(**inputs_dict, return_dict=False)[0]
+
+            # Create modified inputs for second pass
+            inputs_dict_step2 = inputs_dict.copy()
+            if self.cache_input_key in inputs_dict_step2:
+                inputs_dict_step2[self.cache_input_key] = inputs_dict_step2[self.cache_input_key] + torch.randn_like(
+                    inputs_dict_step2[self.cache_input_key]
+                )
+
+            # Second pass - FBC should skip remaining blocks and use cached residuals
+            output_with_cache = model(**inputs_dict_step2, return_dict=False)[0]
+
+        assert output_with_cache is not None, "Model output should not be None with cache enabled."
+        assert not torch.isnan(output_with_cache).any(), "Model output contains NaN with cache enabled."
+
+        # Run same inputs without cache to compare
+        model.disable_cache()
+        output_without_cache = model(**inputs_dict_step2, return_dict=False)[0]
+
+        # Cached output should be different from non-cached output (due to approximation)
+        assert not torch.allclose(output_without_cache, output_with_cache, atol=1e-5), (
+            "Cached output should be different from non-cached output due to cache approximation."
+        )
+
+    @torch.no_grad()
+    def _test_reset_stateful_cache(self):
+        """Test that _reset_stateful_cache resets the FBC cache state (requires cache_context)."""
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+
+        config = self._get_cache_config()
+        model.enable_cache(config)
+
+        with model.cache_context("fbc_test"):
+            _ = model(**inputs_dict, return_dict=False)[0]
+
+        model._reset_stateful_cache()
+
+        model.disable_cache()
+
+    @require_cache_mixin
+    def test_fbc_cache_enable_disable_state(self):
+        self._test_cache_enable_disable_state()
+
+    @require_cache_mixin
+    def test_fbc_cache_double_enable_raises_error(self):
+        self._test_cache_double_enable_raises_error()
+
+    @require_cache_mixin
+    def test_fbc_cache_hooks_registered(self):
+        self._test_cache_hooks_registered()
+
+    @require_cache_mixin
+    def test_fbc_cache_inference(self):
+        self._test_cache_inference()
+
+    @require_cache_mixin
+    def test_fbc_cache_context_manager(self):
+        self._test_cache_context_manager()
+
+    @require_cache_mixin
+    def test_fbc_reset_stateful_cache(self):
+        self._test_reset_stateful_cache()
+
+
+@is_cache
+class FasterCacheConfigMixin:
+    """
+    Base mixin providing FasterCache config.
+
+    Expected class attributes:
+        - model_class: The model class to test (must use CacheMixin)
+    """
+
+    # Default FasterCache config - can be overridden by subclasses
+    FASTER_CACHE_CONFIG = {
+        "spatial_attention_block_skip_range": 2,
+        "spatial_attention_timestep_skip_range": (-1, 901),
+        "tensor_format": "BCHW",
+    }
+
+    def _get_cache_config(self, current_timestep_callback=None):
+        config_kwargs = self.FASTER_CACHE_CONFIG.copy()
+        if current_timestep_callback is None:
+            current_timestep_callback = lambda: 1000  # noqa: E731
+        config_kwargs["current_timestep_callback"] = current_timestep_callback
+        return FasterCacheConfig(**config_kwargs)
+
+    def _get_hook_names(self):
+        return [_FASTER_CACHE_DENOISER_HOOK, _FASTER_CACHE_BLOCK_HOOK]
+
+
+@is_cache
+class FasterCacheTesterMixin(FasterCacheConfigMixin, CacheTesterMixin):
+    """
+    Mixin class for testing FasterCache on models.
+
+    Note: FasterCache is designed for pipeline-level inference with proper CFG batch handling
+    and timestep management. Inference tests are skipped at model level - FasterCache should
+    be tested via pipeline tests (e.g., FluxPipeline, HunyuanVideoPipeline).
+
+    Expected class attributes:
+        - model_class: The model class to test (must use CacheMixin)
+
+    Expected methods to be implemented by subclasses:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: cache
+        Use `pytest -m "not cache"` to skip these tests
+    """
+
+    @torch.no_grad()
+    def _test_cache_inference(self):
+        """Test that model can run inference with FasterCache enabled."""
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+
+        current_timestep = [1000]
+        config = self._get_cache_config(current_timestep_callback=lambda: current_timestep[0])
+
+        model.enable_cache(config)
+
+        # First pass with timestep outside skip range - computes and populates cache
+        current_timestep[0] = 1000
+        _ = model(**inputs_dict, return_dict=False)[0]
+
+        # Move timestep inside skip range so subsequent passes use cache
+        current_timestep[0] = 500
+
+        # Create modified inputs for second pass
+        inputs_dict_step2 = inputs_dict.copy()
+        if self.cache_input_key in inputs_dict_step2:
+            inputs_dict_step2[self.cache_input_key] = inputs_dict_step2[self.cache_input_key] + torch.randn_like(
+                inputs_dict_step2[self.cache_input_key]
+            )
+
+        # Second pass uses cached attention with different inputs
+        output_with_cache = model(**inputs_dict_step2, return_dict=False)[0]
+
+        assert output_with_cache is not None, "Model output should not be None with cache enabled."
+        assert not torch.isnan(output_with_cache).any(), "Model output contains NaN with cache enabled."
+
+        # Run same inputs without cache to compare
+        model.disable_cache()
+        output_without_cache = model(**inputs_dict_step2, return_dict=False)[0]
+
+        # Cached output should be different from non-cached output (due to approximation)
+        assert not torch.allclose(output_without_cache, output_with_cache, atol=1e-5), (
+            "Cached output should be different from non-cached output due to cache approximation."
+        )
+
+    @torch.no_grad()
+    def _test_reset_stateful_cache(self):
+        """Test that _reset_stateful_cache resets the FasterCache state."""
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+
+        config = self._get_cache_config()
+        model.enable_cache(config)
+
+        _ = model(**inputs_dict, return_dict=False)[0]
+
+        model._reset_stateful_cache()
+
+        model.disable_cache()
+
+    @require_cache_mixin
+    def test_faster_cache_enable_disable_state(self):
+        self._test_cache_enable_disable_state()
+
+    @require_cache_mixin
+    def test_faster_cache_double_enable_raises_error(self):
+        self._test_cache_double_enable_raises_error()
+
+    @require_cache_mixin
+    def test_faster_cache_hooks_registered(self):
+        self._test_cache_hooks_registered()
+
+    @require_cache_mixin
+    def test_faster_cache_inference(self):
+        self._test_cache_inference()
+
+    @require_cache_mixin
+    def test_faster_cache_context_manager(self):
+        self._test_cache_context_manager()
+
+    @require_cache_mixin
+    def test_faster_cache_reset_stateful_cache(self):
+        self._test_reset_stateful_cache()

tests/models/testing_utils/common.py

@@ -0,0 +1,666 @@
+# coding=utf-8
+# Copyright 2025 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 json
+import os
+from collections import defaultdict
+from typing import Any, Dict, Optional, Type
+
+import pytest
+import torch
+import torch.nn as nn
+from accelerate.utils.modeling import _get_proper_dtype, compute_module_sizes, dtype_byte_size
+
+from diffusers.utils import SAFE_WEIGHTS_INDEX_NAME, _add_variant, logging
+from diffusers.utils.testing_utils import require_accelerator, require_torch_multi_accelerator
+
+from ...testing_utils import assert_tensors_close, torch_device
+
+
+def named_persistent_module_tensors(
+    module: nn.Module,
+    recurse: bool = False,
+):
+    """
+    A helper function that gathers all the tensors (parameters + persistent buffers) of a given module.
+
+    Args:
+        module (`torch.nn.Module`):
+            The module we want the tensors on.
+        recurse (`bool`, *optional`, defaults to `False`):
+            Whether or not to go look in every submodule or just return the direct parameters and buffers.
+    """
+    yield from module.named_parameters(recurse=recurse)
+
+    for named_buffer in module.named_buffers(recurse=recurse):
+        name, _ = named_buffer
+        # Get parent by splitting on dots and traversing the model
+        parent = module
+        if "." in name:
+            parent_name = name.rsplit(".", 1)[0]
+            for part in parent_name.split("."):
+                parent = getattr(parent, part)
+            name = name.split(".")[-1]
+        if name not in parent._non_persistent_buffers_set:
+            yield named_buffer
+
+
+def compute_module_persistent_sizes(
+    model: nn.Module,
+    dtype: str | torch.device | None = None,
+    special_dtypes: dict[str, str | torch.device] | None = None,
+):
+    """
+    Compute the size of each submodule of a given model (parameters + persistent buffers).
+    """
+    if dtype is not None:
+        dtype = _get_proper_dtype(dtype)
+        dtype_size = dtype_byte_size(dtype)
+    if special_dtypes is not None:
+        special_dtypes = {key: _get_proper_dtype(dtyp) for key, dtyp in special_dtypes.items()}
+        special_dtypes_size = {key: dtype_byte_size(dtyp) for key, dtyp in special_dtypes.items()}
+    module_sizes = defaultdict(int)
+
+    module_list = []
+
+    module_list = named_persistent_module_tensors(model, recurse=True)
+
+    for name, tensor in module_list:
+        if special_dtypes is not None and name in special_dtypes:
+            size = tensor.numel() * special_dtypes_size[name]
+        elif dtype is None:
+            size = tensor.numel() * dtype_byte_size(tensor.dtype)
+        elif str(tensor.dtype).startswith(("torch.uint", "torch.int", "torch.bool")):
+            # According to the code in set_module_tensor_to_device, these types won't be converted
+            # so use their original size here
+            size = tensor.numel() * dtype_byte_size(tensor.dtype)
+        else:
+            size = tensor.numel() * min(dtype_size, dtype_byte_size(tensor.dtype))
+        name_parts = name.split(".")
+        for idx in range(len(name_parts) + 1):
+            module_sizes[".".join(name_parts[:idx])] += size
+
+    return module_sizes
+
+
+def calculate_expected_num_shards(index_map_path):
+    """
+    Calculate expected number of shards from index file.
+
+    Args:
+        index_map_path: Path to the sharded checkpoint index file
+
+    Returns:
+        int: Expected number of shards
+    """
+    with open(index_map_path) as f:
+        weight_map_dict = json.load(f)["weight_map"]
+    first_key = list(weight_map_dict.keys())[0]
+    weight_loc = weight_map_dict[first_key]  # e.g., diffusion_pytorch_model-00001-of-00002.safetensors
+    expected_num_shards = int(weight_loc.split("-")[-1].split(".")[0])
+    return expected_num_shards
+
+
+def check_device_map_is_respected(model, device_map):
+    for param_name, param in model.named_parameters():
+        # Find device in device_map
+        while len(param_name) > 0 and param_name not in device_map:
+            param_name = ".".join(param_name.split(".")[:-1])
+        if param_name not in device_map:
+            raise ValueError("device map is incomplete, it does not contain any device for `param_name`.")
+
+        param_device = device_map[param_name]
+        if param_device in ["cpu", "disk"]:
+            assert param.device == torch.device("meta"), f"Expected device 'meta' for {param_name}, got {param.device}"
+        else:
+            assert param.device == torch.device(param_device), (
+                f"Expected device {param_device} for {param_name}, got {param.device}"
+            )
+
+
+def cast_inputs_to_dtype(inputs, current_dtype, target_dtype):
+    if torch.is_tensor(inputs):
+        return inputs.to(target_dtype) if inputs.dtype == current_dtype else inputs
+    if isinstance(inputs, dict):
+        return {k: cast_inputs_to_dtype(v, current_dtype, target_dtype) for k, v in inputs.items()}
+    if isinstance(inputs, list):
+        return [cast_inputs_to_dtype(v, current_dtype, target_dtype) for v in inputs]
+
+    return inputs
+
+
+class BaseModelTesterConfig:
+    """
+    Base class defining the configuration interface for model testing.
+
+    This class defines the contract that all model test classes must implement.
+    It provides a consistent interface for accessing model configuration, initialization
+    parameters, and test inputs across all testing mixins.
+
+    Required properties (must be implemented by subclasses):
+        - model_class: The model class to test
+
+    Optional properties (can be overridden, have sensible defaults):
+        - pretrained_model_name_or_path: Hub repository ID for pretrained model (default: None)
+        - pretrained_model_kwargs: Additional kwargs for from_pretrained (default: {})
+        - output_shape: Expected output shape for output validation tests (default: None)
+        - model_split_percents: Percentages for model parallelism tests (default: [0.5, 0.7])
+
+    Required methods (must be implemented by subclasses):
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Example usage:
+        class MyModelTestConfig(BaseModelTesterConfig):
+            @property
+            def model_class(self):
+                return MyModel
+
+            @property
+            def pretrained_model_name_or_path(self):
+                return "org/my-model"
+
+            @property
+            def output_shape(self):
+                return (1, 3, 32, 32)
+
+            def get_init_dict(self):
+                return {"in_channels": 3, "out_channels": 3}
+
+            def get_dummy_inputs(self):
+                return {"sample": torch.randn(1, 3, 32, 32, device=torch_device)}
+
+        class TestMyModel(MyModelTestConfig, ModelTesterMixin, QuantizationTesterMixin):
+            pass
+    """
+
+    # ==================== Required Properties ====================
+
+    @property
+    def model_class(self) -> Type[nn.Module]:
+        """The model class to test. Must be implemented by subclasses."""
+        raise NotImplementedError("Subclasses must implement the `model_class` property.")
+
+    # ==================== Optional Properties ====================
+
+    @property
+    def pretrained_model_name_or_path(self) -> Optional[str]:
+        """Hub repository ID for the pretrained model (used for quantization and hub tests)."""
+        return None
+
+    @property
+    def pretrained_model_kwargs(self) -> Dict[str, Any]:
+        """Additional kwargs to pass to from_pretrained (e.g., subfolder, variant)."""
+        return {}
+
+    @property
+    def output_shape(self) -> Optional[tuple]:
+        """Expected output shape for output validation tests."""
+        return None
+
+    @property
+    def model_split_percents(self) -> list:
+        """Percentages for model parallelism tests."""
+        return [0.5, 0.7]
+
+    # ==================== Required Methods ====================
+
+    def get_init_dict(self) -> Dict[str, Any]:
+        """
+        Returns dict of arguments to initialize the model.
+
+        Returns:
+            Dict[str, Any]: Initialization arguments for the model constructor.
+
+        Example:
+            return {
+                "in_channels": 3,
+                "out_channels": 3,
+                "sample_size": 32,
+            }
+        """
+        raise NotImplementedError("Subclasses must implement `get_init_dict()`.")
+
+    def get_dummy_inputs(self) -> Dict[str, Any]:
+        """
+        Returns dict of inputs to pass to the model forward pass.
+
+        Returns:
+            Dict[str, Any]: Input tensors/values for model.forward().
+
+        Example:
+            return {
+                "sample": torch.randn(1, 3, 32, 32, device=torch_device),
+                "timestep": torch.tensor([1], device=torch_device),
+            }
+        """
+        raise NotImplementedError("Subclasses must implement `get_dummy_inputs()`.")
+
+
+class ModelTesterMixin:
+    """
+    Base mixin class for model testing with common test methods.
+
+    This mixin expects the test class to also inherit from BaseModelTesterConfig
+    (or implement its interface) which provides:
+        - model_class: The model class to test
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Example:
+        class MyModelTestConfig(BaseModelTesterConfig):
+            model_class = MyModel
+            def get_init_dict(self): ...
+            def get_dummy_inputs(self): ...
+
+        class TestMyModel(MyModelTestConfig, ModelTesterMixin):
+            pass
+    """
+
+    @torch.no_grad()
+    def test_from_save_pretrained(self, tmp_path, atol=5e-5, rtol=5e-5):
+        torch.manual_seed(0)
+        model = self.model_class(**self.get_init_dict())
+        model.to(torch_device)
+        model.eval()
+
+        model.save_pretrained(tmp_path)
+        new_model = self.model_class.from_pretrained(tmp_path)
+        new_model.to(torch_device)
+
+        for param_name in model.state_dict().keys():
+            param_1 = model.state_dict()[param_name]
+            param_2 = new_model.state_dict()[param_name]
+            assert param_1.shape == param_2.shape, (
+                f"Parameter shape mismatch for {param_name}. Original: {param_1.shape}, loaded: {param_2.shape}"
+            )
+
+        image = model(**self.get_dummy_inputs(), return_dict=False)[0]
+        new_image = new_model(**self.get_dummy_inputs(), return_dict=False)[0]
+
+        assert_tensors_close(image, new_image, atol=atol, rtol=rtol, msg="Models give different forward passes.")
+
+    @torch.no_grad()
+    def test_from_save_pretrained_variant(self, tmp_path, atol=5e-5, rtol=0):
+        model = self.model_class(**self.get_init_dict())
+        model.to(torch_device)
+        model.eval()
+
+        model.save_pretrained(tmp_path, variant="fp16")
+        new_model = self.model_class.from_pretrained(tmp_path, variant="fp16")
+
+        with pytest.raises(OSError) as exc_info:
+            self.model_class.from_pretrained(tmp_path)
+
+        assert "Error no file named diffusion_pytorch_model.bin found in directory" in str(exc_info.value)
+
+        new_model.to(torch_device)
+
+        image = model(**self.get_dummy_inputs(), return_dict=False)[0]
+        new_image = new_model(**self.get_dummy_inputs(), return_dict=False)[0]
+
+        assert_tensors_close(image, new_image, atol=atol, rtol=rtol, msg="Models give different forward passes.")
+
+    @pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16], ids=["fp32", "fp16", "bf16"])
+    def test_from_save_pretrained_dtype(self, tmp_path, dtype):
+        model = self.model_class(**self.get_init_dict())
+        model.to(torch_device)
+        model.eval()
+
+        if torch_device == "mps" and dtype == torch.bfloat16:
+            pytest.skip(reason=f"{dtype} is not supported on {torch_device}")
+
+        model.to(dtype)
+        model.save_pretrained(tmp_path)
+        new_model = self.model_class.from_pretrained(tmp_path, low_cpu_mem_usage=True, torch_dtype=dtype)
+        assert new_model.dtype == dtype
+        if hasattr(self.model_class, "_keep_in_fp32_modules") and self.model_class._keep_in_fp32_modules is None:
+            # When loading without accelerate dtype == torch.float32 if _keep_in_fp32_modules is not None
+            new_model = self.model_class.from_pretrained(tmp_path, low_cpu_mem_usage=False, torch_dtype=dtype)
+            assert new_model.dtype == dtype
+
+    @torch.no_grad()
+    def test_determinism(self, atol=1e-5, rtol=0):
+        model = self.model_class(**self.get_init_dict())
+        model.to(torch_device)
+        model.eval()
+
+        first = model(**self.get_dummy_inputs(), return_dict=False)[0]
+        second = model(**self.get_dummy_inputs(), return_dict=False)[0]
+
+        first_flat = first.flatten()
+        second_flat = second.flatten()
+        mask = ~(torch.isnan(first_flat) | torch.isnan(second_flat))
+        first_filtered = first_flat[mask]
+        second_filtered = second_flat[mask]
+
+        assert_tensors_close(
+            first_filtered, second_filtered, atol=atol, rtol=rtol, msg="Model outputs are not deterministic"
+        )
+
+    @torch.no_grad()
+    def test_output(self, expected_output_shape=None):
+        model = self.model_class(**self.get_init_dict())
+        model.to(torch_device)
+        model.eval()
+
+        inputs_dict = self.get_dummy_inputs()
+        output = model(**inputs_dict, return_dict=False)[0]
+
+        assert output is not None, "Model output is None"
+        assert output[0].shape == expected_output_shape or self.output_shape, (
+            f"Output shape does not match expected. Expected {expected_output_shape}, got {output.shape}"
+        )
+
+    @torch.no_grad()
+    def test_outputs_equivalence(self, atol=1e-5, rtol=0):
+        def set_nan_tensor_to_zero(t):
+            device = t.device
+            if device.type == "mps":
+                t = t.to("cpu")
+            t[t != t] = 0
+            return t.to(device)
+
+        def recursive_check(tuple_object, dict_object):
+            if isinstance(tuple_object, (list, tuple)):
+                for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object.values()):
+                    recursive_check(tuple_iterable_value, dict_iterable_value)
+            elif isinstance(tuple_object, dict):
+                for tuple_iterable_value, dict_iterable_value in zip(tuple_object.values(), dict_object.values()):
+                    recursive_check(tuple_iterable_value, dict_iterable_value)
+            elif tuple_object is None:
+                return
+            else:
+                assert_tensors_close(
+                    set_nan_tensor_to_zero(tuple_object),
+                    set_nan_tensor_to_zero(dict_object),
+                    atol=atol,
+                    rtol=rtol,
+                    msg="Tuple and dict output are not equal",
+                )
+
+        model = self.model_class(**self.get_init_dict())
+        model.to(torch_device)
+        model.eval()
+
+        outputs_dict = model(**self.get_dummy_inputs())
+        outputs_tuple = model(**self.get_dummy_inputs(), return_dict=False)
+
+        recursive_check(outputs_tuple, outputs_dict)
+
+    def test_getattr_is_correct(self, caplog):
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict)
+
+        # save some things to test
+        model.dummy_attribute = 5
+        model.register_to_config(test_attribute=5)
+
+        logger_name = "diffusers.models.modeling_utils"
+        with caplog.at_level(logging.WARNING, logger=logger_name):
+            caplog.clear()
+            assert hasattr(model, "dummy_attribute")
+            assert getattr(model, "dummy_attribute") == 5
+            assert model.dummy_attribute == 5
+
+        # no warning should be thrown
+        assert caplog.text == ""
+
+        with caplog.at_level(logging.WARNING, logger=logger_name):
+            caplog.clear()
+            assert hasattr(model, "save_pretrained")
+            fn = model.save_pretrained
+            fn_1 = getattr(model, "save_pretrained")
+
+            assert fn == fn_1
+
+        # no warning should be thrown
+        assert caplog.text == ""
+
+        # warning should be thrown for config attributes accessed directly
+        with pytest.warns(FutureWarning):
+            assert model.test_attribute == 5
+
+        with pytest.warns(FutureWarning):
+            assert getattr(model, "test_attribute") == 5
+
+        with pytest.raises(AttributeError) as error:
+            model.does_not_exist
+
+        assert str(error.value) == f"'{type(model).__name__}' object has no attribute 'does_not_exist'"
+
+    @require_accelerator
+    @pytest.mark.skipif(
+        torch_device not in ["cuda", "xpu"],
+        reason="float16 and bfloat16 can only be used with an accelerator",
+    )
+    def test_keep_in_fp32_modules(self):
+        model = self.model_class(**self.get_init_dict())
+        fp32_modules = model._keep_in_fp32_modules
+
+        if fp32_modules is None or len(fp32_modules) == 0:
+            pytest.skip("Model does not have _keep_in_fp32_modules defined.")
+
+        # Test with float16
+        model.to(torch_device)
+        model.to(torch.float16)
+
+        for name, param in model.named_parameters():
+            if any(module_to_keep_in_fp32 in name.split(".") for module_to_keep_in_fp32 in fp32_modules):
+                assert param.dtype == torch.float32, f"Parameter {name} should be float32 but got {param.dtype}"
+            else:
+                assert param.dtype == torch.float16, f"Parameter {name} should be float16 but got {param.dtype}"
+
+    @require_accelerator
+    @pytest.mark.skipif(
+        torch_device not in ["cuda", "xpu"],
+        reason="float16 and bfloat16 can only be use for inference with an accelerator",
+    )
+    @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16], ids=["fp16", "bf16"])
+    @torch.no_grad()
+    def test_from_save_pretrained_dtype_inference(self, tmp_path, dtype):
+        model = self.model_class(**self.get_init_dict())
+        model.to(torch_device)
+        fp32_modules = model._keep_in_fp32_modules or []
+
+        model.to(dtype).save_pretrained(tmp_path)
+        model_loaded = self.model_class.from_pretrained(tmp_path, torch_dtype=dtype).to(torch_device)
+
+        for name, param in model_loaded.named_parameters():
+            if fp32_modules and any(
+                module_to_keep_in_fp32 in name.split(".") for module_to_keep_in_fp32 in fp32_modules
+            ):
+                assert param.data.dtype == torch.float32
+            else:
+                assert param.data.dtype == dtype
+
+        inputs = cast_inputs_to_dtype(self.get_dummy_inputs(), torch.float32, dtype)
+        output = model(**inputs, return_dict=False)[0]
+        output_loaded = model_loaded(**inputs, return_dict=False)[0]
+
+        self._check_dtype_inference_output(output, output_loaded, dtype)
+
+    def _check_dtype_inference_output(self, output, output_loaded, dtype, atol=1e-4, rtol=0):
+        """Check dtype inference output with configurable tolerance."""
+        assert_tensors_close(
+            output, output_loaded, atol=atol, rtol=rtol, msg=f"Loaded model output differs for {dtype}"
+        )
+
+    @require_accelerator
+    @torch.no_grad()
+    def test_sharded_checkpoints(self, tmp_path, atol=1e-5, rtol=0):
+        torch.manual_seed(0)
+        config = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**config).eval()
+        model = model.to(torch_device)
+
+        base_output = model(**inputs_dict, return_dict=False)[0]
+
+        model_size = compute_module_persistent_sizes(model)[""]
+        max_shard_size = int((model_size * 0.75) / (2**10))  # Convert to KB as these test models are small
+
+        model.cpu().save_pretrained(tmp_path, max_shard_size=f"{max_shard_size}KB")
+        assert os.path.exists(os.path.join(tmp_path, SAFE_WEIGHTS_INDEX_NAME)), "Index file should exist"
+
+        # Check if the right number of shards exists
+        expected_num_shards = calculate_expected_num_shards(os.path.join(tmp_path, SAFE_WEIGHTS_INDEX_NAME))
+        actual_num_shards = len([file for file in os.listdir(tmp_path) if file.endswith(".safetensors")])
+        assert actual_num_shards == expected_num_shards, (
+            f"Expected {expected_num_shards} shards, got {actual_num_shards}"
+        )
+
+        new_model = self.model_class.from_pretrained(tmp_path).eval()
+        new_model = new_model.to(torch_device)
+
+        torch.manual_seed(0)
+        inputs_dict_new = self.get_dummy_inputs()
+        new_output = new_model(**inputs_dict_new, return_dict=False)[0]
+
+        assert_tensors_close(
+            base_output, new_output, atol=atol, rtol=rtol, msg="Output should match after sharded save/load"
+        )
+
+    @require_accelerator
+    @torch.no_grad()
+    def test_sharded_checkpoints_with_variant(self, tmp_path, atol=1e-5, rtol=0):
+        torch.manual_seed(0)
+        config = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**config).eval()
+        model = model.to(torch_device)
+
+        base_output = model(**inputs_dict, return_dict=False)[0]
+
+        model_size = compute_module_persistent_sizes(model)[""]
+        max_shard_size = int((model_size * 0.75) / (2**10))  # Convert to KB as these test models are small
+        variant = "fp16"
+
+        model.cpu().save_pretrained(tmp_path, max_shard_size=f"{max_shard_size}KB", variant=variant)
+
+        index_filename = _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
+        assert os.path.exists(os.path.join(tmp_path, index_filename)), (
+            f"Variant index file {index_filename} should exist"
+        )
+
+        # Check if the right number of shards exists
+        expected_num_shards = calculate_expected_num_shards(os.path.join(tmp_path, index_filename))
+        actual_num_shards = len([file for file in os.listdir(tmp_path) if file.endswith(".safetensors")])
+        assert actual_num_shards == expected_num_shards, (
+            f"Expected {expected_num_shards} shards, got {actual_num_shards}"
+        )
+
+        new_model = self.model_class.from_pretrained(tmp_path, variant=variant).eval()
+        new_model = new_model.to(torch_device)
+
+        torch.manual_seed(0)
+        inputs_dict_new = self.get_dummy_inputs()
+        new_output = new_model(**inputs_dict_new, return_dict=False)[0]
+
+        assert_tensors_close(
+            base_output, new_output, atol=atol, rtol=rtol, msg="Output should match after variant sharded save/load"
+        )
+
+    @torch.no_grad()
+    def test_sharded_checkpoints_with_parallel_loading(self, tmp_path, atol=1e-5, rtol=0):
+        from diffusers.utils import constants
+
+        torch.manual_seed(0)
+        config = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**config).eval()
+        model = model.to(torch_device)
+
+        base_output = model(**inputs_dict, return_dict=False)[0]
+
+        model_size = compute_module_persistent_sizes(model)[""]
+        max_shard_size = int((model_size * 0.75) / (2**10))  # Convert to KB as these test models are small
+
+        # Save original values to restore after test
+        original_parallel_loading = constants.HF_ENABLE_PARALLEL_LOADING
+        original_parallel_workers = getattr(constants, "HF_PARALLEL_WORKERS", None)
+
+        try:
+            model.cpu().save_pretrained(tmp_path, max_shard_size=f"{max_shard_size}KB")
+            assert os.path.exists(os.path.join(tmp_path, SAFE_WEIGHTS_INDEX_NAME)), "Index file should exist"
+
+            # Check if the right number of shards exists
+            expected_num_shards = calculate_expected_num_shards(os.path.join(tmp_path, SAFE_WEIGHTS_INDEX_NAME))
+            actual_num_shards = len([file for file in os.listdir(tmp_path) if file.endswith(".safetensors")])
+            assert actual_num_shards == expected_num_shards, (
+                f"Expected {expected_num_shards} shards, got {actual_num_shards}"
+            )
+
+            # Load without parallel loading
+            constants.HF_ENABLE_PARALLEL_LOADING = False
+            model_sequential = self.model_class.from_pretrained(tmp_path).eval()
+            model_sequential = model_sequential.to(torch_device)
+
+            # Load with parallel loading
+            constants.HF_ENABLE_PARALLEL_LOADING = True
+            constants.DEFAULT_HF_PARALLEL_LOADING_WORKERS = 2
+
+            torch.manual_seed(0)
+            model_parallel = self.model_class.from_pretrained(tmp_path).eval()
+            model_parallel = model_parallel.to(torch_device)
+
+            torch.manual_seed(0)
+            inputs_dict_parallel = self.get_dummy_inputs()
+            output_parallel = model_parallel(**inputs_dict_parallel, return_dict=False)[0]
+
+            assert_tensors_close(
+                base_output, output_parallel, atol=atol, rtol=rtol, msg="Output should match with parallel loading"
+            )
+
+        finally:
+            # Restore original values
+            constants.HF_ENABLE_PARALLEL_LOADING = original_parallel_loading
+            if original_parallel_workers is not None:
+                constants.HF_PARALLEL_WORKERS = original_parallel_workers
+
+    @require_torch_multi_accelerator
+    @torch.no_grad()
+    def test_model_parallelism(self, tmp_path, atol=1e-5, rtol=0):
+        if self.model_class._no_split_modules is None:
+            pytest.skip("Test not supported for this model as `_no_split_modules` is not set.")
+
+        config = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**config).eval()
+
+        model = model.to(torch_device)
+
+        torch.manual_seed(0)
+        base_output = model(**inputs_dict, return_dict=False)[0]
+
+        model_size = compute_module_sizes(model)[""]
+        max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents]
+
+        model.cpu().save_pretrained(tmp_path)
+
+        for max_size in max_gpu_sizes:
+            max_memory = {0: max_size, 1: model_size * 2, "cpu": model_size * 2}
+            new_model = self.model_class.from_pretrained(tmp_path, device_map="auto", max_memory=max_memory)
+            # Making sure part of the model will be on GPU 0 and GPU 1
+            assert set(new_model.hf_device_map.values()) == {0, 1}, "Model should be split across GPUs"
+
+            check_device_map_is_respected(new_model, new_model.hf_device_map)
+
+            torch.manual_seed(0)
+            new_output = new_model(**inputs_dict, return_dict=False)[0]
+
+            assert_tensors_close(
+                base_output, new_output, atol=atol, rtol=rtol, msg="Output should match with model parallelism"
+            )

tests/models/testing_utils/compile.py

@@ -0,0 +1,166 @@
+# coding=utf-8
+# Copyright 2025 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 os
+
+import pytest
+import torch
+
+from ...testing_utils import (
+    backend_empty_cache,
+    is_torch_compile,
+    require_accelerator,
+    require_torch_version_greater,
+    torch_device,
+)
+
+
+@is_torch_compile
+@require_accelerator
+@require_torch_version_greater("2.7.1")
+class TorchCompileTesterMixin:
+    """
+    Mixin class for testing torch.compile functionality on models.
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Optional properties:
+        - different_shapes_for_compilation: List of (height, width) tuples for dynamic shape testing (default: None)
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: compile
+        Use `pytest -m "not compile"` to skip these tests
+    """
+
+    @property
+    def different_shapes_for_compilation(self) -> list[tuple[int, int]] | None:
+        """Optional list of (height, width) tuples for dynamic shape testing."""
+        return None
+
+    def setup_method(self):
+        torch.compiler.reset()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def teardown_method(self):
+        torch.compiler.reset()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    @torch.no_grad()
+    def test_torch_compile_recompilation_and_graph_break(self):
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+        model = torch.compile(model, fullgraph=True)
+
+        with (
+            torch._inductor.utils.fresh_inductor_cache(),
+            torch._dynamo.config.patch(error_on_recompile=True),
+        ):
+            _ = model(**inputs_dict)
+            _ = model(**inputs_dict)
+
+    @torch.no_grad()
+    def test_torch_compile_repeated_blocks(self):
+        if self.model_class._repeated_blocks is None:
+            pytest.skip("Skipping test as the model class doesn't have `_repeated_blocks` set.")
+
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+        model.compile_repeated_blocks(fullgraph=True)
+
+        recompile_limit = 1
+        if self.model_class.__name__ == "UNet2DConditionModel":
+            recompile_limit = 2
+
+        with (
+            torch._inductor.utils.fresh_inductor_cache(),
+            torch._dynamo.config.patch(recompile_limit=recompile_limit),
+        ):
+            _ = model(**inputs_dict)
+            _ = model(**inputs_dict)
+
+    @torch.no_grad()
+    def test_compile_with_group_offloading(self):
+        if not self.model_class._supports_group_offloading:
+            pytest.skip("Model does not support group offloading.")
+
+        torch._dynamo.config.cache_size_limit = 10000
+
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict)
+        model.eval()
+
+        group_offload_kwargs = {
+            "onload_device": torch_device,
+            "offload_device": "cpu",
+            "offload_type": "block_level",
+            "num_blocks_per_group": 1,
+            "use_stream": True,
+            "non_blocking": True,
+        }
+        model.enable_group_offload(**group_offload_kwargs)
+        model.compile()
+
+        _ = model(**inputs_dict)
+        _ = model(**inputs_dict)
+
+    @torch.no_grad()
+    def test_compile_on_different_shapes(self):
+        if self.different_shapes_for_compilation is None:
+            pytest.skip(f"Skipping as `different_shapes_for_compilation` is not set for {self.__class__.__name__}.")
+        torch.fx.experimental._config.use_duck_shape = False
+
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+        model = torch.compile(model, fullgraph=True, dynamic=True)
+
+        for height, width in self.different_shapes_for_compilation:
+            with torch._dynamo.config.patch(error_on_recompile=True):
+                inputs_dict = self.get_dummy_inputs(height=height, width=width)
+                _ = model(**inputs_dict)
+
+    @torch.no_grad()
+    def test_compile_works_with_aot(self, tmp_path):
+        from torch._inductor.package import load_package
+
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+
+        model = self.model_class(**init_dict).to(torch_device)
+        exported_model = torch.export.export(model, args=(), kwargs=inputs_dict)
+
+        package_path = os.path.join(str(tmp_path), f"{self.model_class.__name__}.pt2")
+        _ = torch._inductor.aoti_compile_and_package(exported_model, package_path=package_path)
+        assert os.path.exists(package_path), f"Package file not created at {package_path}"
+        loaded_binary = load_package(package_path, run_single_threaded=True)
+
+        model.forward = loaded_binary
+
+        _ = model(**inputs_dict)
+        _ = model(**inputs_dict)

tests/models/testing_utils/ip_adapter.py

@@ -0,0 +1,158 @@
+# coding=utf-8
+# Copyright 2025 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 pytest
+import torch
+
+from ...testing_utils import backend_empty_cache, is_ip_adapter, torch_device
+
+
+def check_if_ip_adapter_correctly_set(model, processor_cls) -> bool:
+    """
+    Check if IP Adapter processors are correctly set in the model.
+
+    Args:
+        model: The model to check
+
+    Returns:
+        bool: True if IP Adapter is correctly set, False otherwise
+    """
+    for module in model.attn_processors.values():
+        if isinstance(module, processor_cls):
+            return True
+    return False
+
+
+@is_ip_adapter
+class IPAdapterTesterMixin:
+    """
+    Mixin class for testing IP Adapter functionality on models.
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Required properties (must be implemented by subclasses):
+        - ip_adapter_processor_cls: The IP Adapter processor class to use
+
+    Required methods (must be implemented by subclasses):
+        - create_ip_adapter_state_dict(): Creates IP Adapter state dict for testing
+        - modify_inputs_for_ip_adapter(): Modifies inputs to include IP Adapter data
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: ip_adapter
+        Use `pytest -m "not ip_adapter"` to skip these tests
+    """
+
+    def setup_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def teardown_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    @property
+    def ip_adapter_processor_cls(self):
+        """IP Adapter processor class to use for testing. Must be implemented by subclasses."""
+        raise NotImplementedError("Subclasses must implement the `ip_adapter_processor_cls` property.")
+
+    def create_ip_adapter_state_dict(self, model):
+        raise NotImplementedError("child class must implement method to create IPAdapter State Dict")
+
+    def modify_inputs_for_ip_adapter(self, model, inputs_dict):
+        raise NotImplementedError("child class must implement method to create IPAdapter model inputs")
+
+    @torch.no_grad()
+    def test_load_ip_adapter(self):
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        torch.manual_seed(0)
+        output_no_adapter = model(**inputs_dict, return_dict=False)[0]
+
+        ip_adapter_state_dict = self.create_ip_adapter_state_dict(model)
+
+        model._load_ip_adapter_weights([ip_adapter_state_dict])
+        assert check_if_ip_adapter_correctly_set(model, self.ip_adapter_processor_cls), (
+            "IP Adapter processors not set correctly"
+        )
+
+        inputs_dict_with_adapter = self.modify_inputs_for_ip_adapter(model, inputs_dict.copy())
+        outputs_with_adapter = model(**inputs_dict_with_adapter, return_dict=False)[0]
+
+        assert not torch.allclose(output_no_adapter, outputs_with_adapter, atol=1e-4, rtol=1e-4), (
+            "Output should differ with IP Adapter enabled"
+        )
+
+    @pytest.mark.skip(
+        reason="Setting IP Adapter scale is not defined at the model level. Enable this test after refactoring"
+    )
+    def test_ip_adapter_scale(self):
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        ip_adapter_state_dict = self.create_ip_adapter_state_dict(model)
+        model._load_ip_adapter_weights([ip_adapter_state_dict])
+
+        inputs_dict_with_adapter = self.modify_inputs_for_ip_adapter(model, inputs_dict.copy())
+
+        # Test scale = 0.0 (no effect)
+        model.set_ip_adapter_scale(0.0)
+        torch.manual_seed(0)
+        output_scale_zero = model(**inputs_dict_with_adapter, return_dict=False)[0]
+
+        # Test scale = 1.0 (full effect)
+        model.set_ip_adapter_scale(1.0)
+        torch.manual_seed(0)
+        output_scale_one = model(**inputs_dict_with_adapter, return_dict=False)[0]
+
+        # Outputs should differ with different scales
+        assert not torch.allclose(output_scale_zero, output_scale_one, atol=1e-4, rtol=1e-4), (
+            "Output should differ with different IP Adapter scales"
+        )
+
+    @pytest.mark.skip(
+        reason="Unloading IP Adapter is not defined at the model level. Enable this test after refactoring"
+    )
+    def test_unload_ip_adapter(self):
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        # Save original processors
+        original_processors = {k: type(v).__name__ for k, v in model.attn_processors.items()}
+
+        # Create and load IP adapter
+        ip_adapter_state_dict = self.create_ip_adapter_state_dict(model)
+        model._load_ip_adapter_weights([ip_adapter_state_dict])
+
+        assert check_if_ip_adapter_correctly_set(model, self.ip_adapter_processor_cls), "IP Adapter should be set"
+
+        # Unload IP adapter
+        model.unload_ip_adapter()
+
+        assert not check_if_ip_adapter_correctly_set(model, self.ip_adapter_processor_cls), (
+            "IP Adapter should be unloaded"
+        )
+
+        # Verify processors are restored
+        current_processors = {k: type(v).__name__ for k, v in model.attn_processors.items()}
+        assert original_processors == current_processors, "Processors should be restored after unload"

tests/models/testing_utils/lora.py

@@ -0,0 +1,555 @@
+# coding=utf-8
+# Copyright 2025 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 json
+import os
+import re
+
+import pytest
+import safetensors.torch
+import torch
+import torch.nn as nn
+
+from diffusers.utils.import_utils import is_peft_available
+from diffusers.utils.testing_utils import check_if_dicts_are_equal
+
+from ...testing_utils import (
+    assert_tensors_close,
+    backend_empty_cache,
+    is_lora,
+    is_torch_compile,
+    require_peft_backend,
+    require_peft_version_greater,
+    require_torch_accelerator,
+    require_torch_version_greater,
+    torch_device,
+)
+
+
+if is_peft_available():
+    from diffusers.loaders.peft import PeftAdapterMixin
+
+
+def check_if_lora_correctly_set(model) -> bool:
+    """
+    Check if LoRA layers are correctly set in the model.
+
+    Args:
+        model: The model to check
+
+    Returns:
+        bool: True if LoRA is correctly set, False otherwise
+    """
+    from peft.tuners.tuners_utils import BaseTunerLayer
+
+    for module in model.modules():
+        if isinstance(module, BaseTunerLayer):
+            return True
+    return False
+
+
+@is_lora
+@require_peft_backend
+class LoraTesterMixin:
+    """
+    Mixin class for testing LoRA/PEFT functionality on models.
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: lora
+        Use `pytest -m "not lora"` to skip these tests
+    """
+
+    def setup_method(self):
+        if not issubclass(self.model_class, PeftAdapterMixin):
+            pytest.skip(f"PEFT is not supported for this model ({self.model_class.__name__}).")
+
+    @torch.no_grad()
+    def test_save_load_lora_adapter(self, tmp_path, rank=4, lora_alpha=4, use_dora=False, atol=1e-4, rtol=1e-4):
+        from peft import LoraConfig
+        from peft.utils import get_peft_model_state_dict
+
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        torch.manual_seed(0)
+        output_no_lora = model(**inputs_dict, return_dict=False)[0]
+
+        denoiser_lora_config = LoraConfig(
+            r=rank,
+            lora_alpha=lora_alpha,
+            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
+            init_lora_weights=False,
+            use_dora=use_dora,
+        )
+        model.add_adapter(denoiser_lora_config)
+        assert check_if_lora_correctly_set(model), "LoRA layers not set correctly"
+
+        torch.manual_seed(0)
+        outputs_with_lora = model(**inputs_dict, return_dict=False)[0]
+
+        assert not torch.allclose(output_no_lora, outputs_with_lora, atol=atol, rtol=rtol), (
+            "Output should differ with LoRA enabled"
+        )
+
+        model.save_lora_adapter(tmp_path)
+        assert os.path.isfile(os.path.join(tmp_path, "pytorch_lora_weights.safetensors")), (
+            "LoRA weights file not created"
+        )
+
+        state_dict_loaded = safetensors.torch.load_file(os.path.join(tmp_path, "pytorch_lora_weights.safetensors"))
+
+        model.unload_lora()
+        assert not check_if_lora_correctly_set(model), "LoRA should be unloaded"
+
+        model.load_lora_adapter(tmp_path, prefix=None, use_safetensors=True)
+        state_dict_retrieved = get_peft_model_state_dict(model, adapter_name="default_0")
+
+        for k in state_dict_loaded:
+            loaded_v = state_dict_loaded[k]
+            retrieved_v = state_dict_retrieved[k].to(loaded_v.device)
+            assert_tensors_close(loaded_v, retrieved_v, atol=atol, rtol=rtol, msg=f"Mismatch in LoRA weight {k}")
+
+        assert check_if_lora_correctly_set(model), "LoRA layers not set correctly after reload"
+
+        torch.manual_seed(0)
+        outputs_with_lora_2 = model(**inputs_dict, return_dict=False)[0]
+
+        assert not torch.allclose(output_no_lora, outputs_with_lora_2, atol=atol, rtol=rtol), (
+            "Output should differ with LoRA enabled"
+        )
+        assert_tensors_close(
+            outputs_with_lora,
+            outputs_with_lora_2,
+            atol=atol,
+            rtol=rtol,
+            msg="Outputs should match before and after save/load",
+        )
+
+    def test_lora_wrong_adapter_name_raises_error(self, tmp_path):
+        from peft import LoraConfig
+
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        denoiser_lora_config = LoraConfig(
+            r=4,
+            lora_alpha=4,
+            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
+            init_lora_weights=False,
+            use_dora=False,
+        )
+        model.add_adapter(denoiser_lora_config)
+        assert check_if_lora_correctly_set(model), "LoRA layers not set correctly"
+
+        wrong_name = "foo"
+        with pytest.raises(ValueError) as exc_info:
+            model.save_lora_adapter(tmp_path, adapter_name=wrong_name)
+
+        assert f"Adapter name {wrong_name} not found in the model." in str(exc_info.value)
+
+    def test_lora_adapter_metadata_is_loaded_correctly(self, tmp_path, rank=4, lora_alpha=4, use_dora=False):
+        from peft import LoraConfig
+
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        denoiser_lora_config = LoraConfig(
+            r=rank,
+            lora_alpha=lora_alpha,
+            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
+            init_lora_weights=False,
+            use_dora=use_dora,
+        )
+        model.add_adapter(denoiser_lora_config)
+        metadata = model.peft_config["default"].to_dict()
+        assert check_if_lora_correctly_set(model), "LoRA layers not set correctly"
+
+        model.save_lora_adapter(tmp_path)
+        model_file = os.path.join(tmp_path, "pytorch_lora_weights.safetensors")
+        assert os.path.isfile(model_file), "LoRA weights file not created"
+
+        model.unload_lora()
+        assert not check_if_lora_correctly_set(model), "LoRA should be unloaded"
+
+        model.load_lora_adapter(tmp_path, prefix=None, use_safetensors=True)
+        parsed_metadata = model.peft_config["default_0"].to_dict()
+        check_if_dicts_are_equal(metadata, parsed_metadata)
+
+    def test_lora_adapter_wrong_metadata_raises_error(self, tmp_path):
+        from peft import LoraConfig
+
+        from diffusers.loaders.lora_base import LORA_ADAPTER_METADATA_KEY
+
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        denoiser_lora_config = LoraConfig(
+            r=4,
+            lora_alpha=4,
+            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
+            init_lora_weights=False,
+            use_dora=False,
+        )
+        model.add_adapter(denoiser_lora_config)
+        assert check_if_lora_correctly_set(model), "LoRA layers not set correctly"
+
+        model.save_lora_adapter(tmp_path)
+        model_file = os.path.join(tmp_path, "pytorch_lora_weights.safetensors")
+        assert os.path.isfile(model_file), "LoRA weights file not created"
+
+        # Perturb the metadata in the state dict
+        loaded_state_dict = safetensors.torch.load_file(model_file)
+        metadata = {"format": "pt"}
+        lora_adapter_metadata = denoiser_lora_config.to_dict()
+        lora_adapter_metadata.update({"foo": 1, "bar": 2})
+        for key, value in lora_adapter_metadata.items():
+            if isinstance(value, set):
+                lora_adapter_metadata[key] = list(value)
+        metadata[LORA_ADAPTER_METADATA_KEY] = json.dumps(lora_adapter_metadata, indent=2, sort_keys=True)
+        safetensors.torch.save_file(loaded_state_dict, model_file, metadata=metadata)
+
+        model.unload_lora()
+        assert not check_if_lora_correctly_set(model), "LoRA should be unloaded"
+
+        with pytest.raises(TypeError) as exc_info:
+            model.load_lora_adapter(tmp_path, prefix=None, use_safetensors=True)
+        assert "`LoraConfig` class could not be instantiated" in str(exc_info.value)
+
+
+@is_lora
+@is_torch_compile
+@require_peft_backend
+@require_peft_version_greater("0.14.0")
+@require_torch_version_greater("2.7.1")
+@require_torch_accelerator
+class LoraHotSwappingForModelTesterMixin:
+    """
+    Mixin class for testing LoRA hot swapping functionality on models.
+
+    Test that hotswapping does not result in recompilation on the model directly.
+    We're not extensively testing the hotswapping functionality since it is implemented in PEFT
+    and is extensively tested there. The goal of this test is specifically to ensure that
+    hotswapping with diffusers does not require recompilation.
+
+    See https://github.com/huggingface/peft/blob/eaab05e18d51fb4cce20a73c9acd82a00c013b83/tests/test_gpu_examples.py#L4252
+    for the analogous PEFT test.
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Optional properties:
+        - different_shapes_for_compilation: List of (height, width) tuples for dynamic compilation tests (default: None)
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest marks: lora, torch_compile
+        Use `pytest -m "not lora"` or `pytest -m "not torch_compile"` to skip these tests
+    """
+
+    @property
+    def different_shapes_for_compilation(self) -> list[tuple[int, int]] | None:
+        """Optional list of (height, width) tuples for dynamic compilation tests."""
+        return None
+
+    def setup_method(self):
+        if not issubclass(self.model_class, PeftAdapterMixin):
+            pytest.skip(f"PEFT is not supported for this model ({self.model_class.__name__}).")
+
+    def teardown_method(self):
+        # It is critical that the dynamo cache is reset for each test. Otherwise, if the test re-uses the same model,
+        # there will be recompilation errors, as torch caches the model when run in the same process.
+        torch.compiler.reset()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def _get_lora_config(self, lora_rank, lora_alpha, target_modules):
+        from peft import LoraConfig
+
+        lora_config = LoraConfig(
+            r=lora_rank,
+            lora_alpha=lora_alpha,
+            target_modules=target_modules,
+            init_lora_weights=False,
+            use_dora=False,
+        )
+        return lora_config
+
+    def _get_linear_module_name_other_than_attn(self, model):
+        linear_names = [
+            name for name, module in model.named_modules() if isinstance(module, nn.Linear) and "to_" not in name
+        ]
+        return linear_names[0]
+
+    def _check_model_hotswap(
+        self, tmp_path, do_compile, rank0, rank1, target_modules0, target_modules1=None, atol=5e-3, rtol=5e-3
+    ):
+        """
+        Check that hotswapping works on a model.
+
+        Steps:
+        - create 2 LoRA adapters and save them
+        - load the first adapter
+        - hotswap the second adapter
+        - check that the outputs are correct
+        - optionally compile the model
+        - optionally check if recompilations happen on different shapes
+
+        Note: We set rank == alpha here because save_lora_adapter does not save the alpha scalings, thus the test would
+        fail if the values are different. Since rank != alpha does not matter for the purpose of this test, this is
+        fine.
+        """
+        different_shapes = self.different_shapes_for_compilation
+        # create 2 adapters with different ranks and alphas
+        torch.manual_seed(0)
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        alpha0, alpha1 = rank0, rank1
+        max_rank = max([rank0, rank1])
+        if target_modules1 is None:
+            target_modules1 = target_modules0[:]
+        lora_config0 = self._get_lora_config(rank0, alpha0, target_modules0)
+        lora_config1 = self._get_lora_config(rank1, alpha1, target_modules1)
+
+        model.add_adapter(lora_config0, adapter_name="adapter0")
+        with torch.inference_mode():
+            torch.manual_seed(0)
+            output0_before = model(**inputs_dict)["sample"]
+
+        model.add_adapter(lora_config1, adapter_name="adapter1")
+        model.set_adapter("adapter1")
+        with torch.inference_mode():
+            torch.manual_seed(0)
+            output1_before = model(**inputs_dict)["sample"]
+
+        # sanity checks:
+        assert not torch.allclose(output0_before, output1_before, atol=atol, rtol=rtol)
+        assert not (output0_before == 0).all()
+        assert not (output1_before == 0).all()
+
+        # save the adapter checkpoints
+        model.save_lora_adapter(os.path.join(tmp_path, "0"), safe_serialization=True, adapter_name="adapter0")
+        model.save_lora_adapter(os.path.join(tmp_path, "1"), safe_serialization=True, adapter_name="adapter1")
+        del model
+
+        # load the first adapter
+        torch.manual_seed(0)
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+
+        if do_compile or (rank0 != rank1):
+            # no need to prepare if the model is not compiled or if the ranks are identical
+            model.enable_lora_hotswap(target_rank=max_rank)
+
+        file_name0 = os.path.join(os.path.join(tmp_path, "0"), "pytorch_lora_weights.safetensors")
+        file_name1 = os.path.join(os.path.join(tmp_path, "1"), "pytorch_lora_weights.safetensors")
+        model.load_lora_adapter(file_name0, safe_serialization=True, adapter_name="adapter0", prefix=None)
+
+        if do_compile:
+            model = torch.compile(model, mode="reduce-overhead", dynamic=different_shapes is not None)
+
+        with torch.inference_mode():
+            # additionally check if dynamic compilation works.
+            if different_shapes is not None:
+                for height, width in different_shapes:
+                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
+                    _ = model(**new_inputs_dict)
+            else:
+                output0_after = model(**inputs_dict)["sample"]
+                assert_tensors_close(
+                    output0_before, output0_after, atol=atol, rtol=rtol, msg="Output mismatch after loading adapter0"
+                )
+
+        # hotswap the 2nd adapter
+        model.load_lora_adapter(file_name1, adapter_name="adapter0", hotswap=True, prefix=None)
+
+        # we need to call forward to potentially trigger recompilation
+        with torch.inference_mode():
+            if different_shapes is not None:
+                for height, width in different_shapes:
+                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
+                    _ = model(**new_inputs_dict)
+            else:
+                output1_after = model(**inputs_dict)["sample"]
+                assert_tensors_close(
+                    output1_before,
+                    output1_after,
+                    atol=atol,
+                    rtol=rtol,
+                    msg="Output mismatch after hotswapping to adapter1",
+                )
+
+        # check error when not passing valid adapter name
+        name = "does-not-exist"
+        msg = f"Trying to hotswap LoRA adapter '{name}' but there is no existing adapter by that name"
+        with pytest.raises(ValueError, match=re.escape(msg)):
+            model.load_lora_adapter(file_name1, adapter_name=name, hotswap=True, prefix=None)
+
+    @pytest.mark.parametrize("rank0,rank1", [(11, 11), (7, 13), (13, 7)])
+    def test_hotswapping_model(self, tmp_path, rank0, rank1):
+        self._check_model_hotswap(
+            tmp_path, do_compile=False, rank0=rank0, rank1=rank1, target_modules0=["to_q", "to_k", "to_v", "to_out.0"]
+        )
+
+    @pytest.mark.parametrize("rank0,rank1", [(11, 11), (7, 13), (13, 7)])
+    def test_hotswapping_compiled_model_linear(self, tmp_path, rank0, rank1):
+        # It's important to add this context to raise an error on recompilation
+        target_modules = ["to_q", "to_k", "to_v", "to_out.0"]
+        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
+            self._check_model_hotswap(
+                tmp_path, do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules
+            )
+
+    @pytest.mark.parametrize("rank0,rank1", [(11, 11), (7, 13), (13, 7)])
+    def test_hotswapping_compiled_model_conv2d(self, tmp_path, rank0, rank1):
+        if "unet" not in self.model_class.__name__.lower():
+            pytest.skip("Test only applies to UNet.")
+
+        # It's important to add this context to raise an error on recompilation
+        target_modules = ["conv", "conv1", "conv2"]
+        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
+            self._check_model_hotswap(
+                tmp_path, do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules
+            )
+
+    @pytest.mark.parametrize("rank0,rank1", [(11, 11), (7, 13), (13, 7)])
+    def test_hotswapping_compiled_model_both_linear_and_conv2d(self, tmp_path, rank0, rank1):
+        if "unet" not in self.model_class.__name__.lower():
+            pytest.skip("Test only applies to UNet.")
+
+        # It's important to add this context to raise an error on recompilation
+        target_modules = ["to_q", "conv"]
+        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
+            self._check_model_hotswap(
+                tmp_path, do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules
+            )
+
+    @pytest.mark.parametrize("rank0,rank1", [(11, 11), (7, 13), (13, 7)])
+    def test_hotswapping_compiled_model_both_linear_and_other(self, tmp_path, rank0, rank1):
+        # In `test_hotswapping_compiled_model_both_linear_and_conv2d()`, we check if we can do hotswapping
+        # with `torch.compile()` for models that have both linear and conv layers. In this test, we check
+        # if we can target a linear layer from the transformer blocks and another linear layer from non-attention
+        # block.
+        target_modules = ["to_q"]
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict)
+
+        target_modules.append(self._get_linear_module_name_other_than_attn(model))
+        del model
+
+        # It's important to add this context to raise an error on recompilation
+        with torch._dynamo.config.patch(error_on_recompile=True):
+            self._check_model_hotswap(
+                tmp_path, do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules
+            )
+
+    def test_enable_lora_hotswap_called_after_adapter_added_raises(self):
+        # ensure that enable_lora_hotswap is called before loading the first adapter
+        lora_config = self._get_lora_config(8, 8, target_modules=["to_q"])
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.add_adapter(lora_config)
+
+        msg = re.escape("Call `enable_lora_hotswap` before loading the first adapter.")
+        with pytest.raises(RuntimeError, match=msg):
+            model.enable_lora_hotswap(target_rank=32)
+
+    def test_enable_lora_hotswap_called_after_adapter_added_warning(self, caplog):
+        # ensure that enable_lora_hotswap is called before loading the first adapter
+        import logging
+
+        lora_config = self._get_lora_config(8, 8, target_modules=["to_q"])
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.add_adapter(lora_config)
+        msg = (
+            "It is recommended to call `enable_lora_hotswap` before loading the first adapter to avoid recompilation."
+        )
+        with caplog.at_level(logging.WARNING):
+            model.enable_lora_hotswap(target_rank=32, check_compiled="warn")
+            assert any(msg in record.message for record in caplog.records)
+
+    def test_enable_lora_hotswap_called_after_adapter_added_ignore(self, caplog):
+        # check possibility to ignore the error/warning
+        import logging
+
+        lora_config = self._get_lora_config(8, 8, target_modules=["to_q"])
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.add_adapter(lora_config)
+        with caplog.at_level(logging.WARNING):
+            model.enable_lora_hotswap(target_rank=32, check_compiled="ignore")
+            assert len(caplog.records) == 0
+
+    def test_enable_lora_hotswap_wrong_check_compiled_argument_raises(self):
+        # check that wrong argument value raises an error
+        lora_config = self._get_lora_config(8, 8, target_modules=["to_q"])
+        init_dict = self.get_init_dict()
+        model = self.model_class(**init_dict).to(torch_device)
+        model.add_adapter(lora_config)
+        msg = re.escape("check_compiles should be one of 'error', 'warn', or 'ignore', got 'wrong-argument' instead.")
+        with pytest.raises(ValueError, match=msg):
+            model.enable_lora_hotswap(target_rank=32, check_compiled="wrong-argument")
+
+    def test_hotswap_second_adapter_targets_more_layers_raises(self, tmp_path, caplog):
+        # check the error and log
+        import logging
+
+        # at the moment, PEFT requires the 2nd adapter to target the same or a subset of layers
+        target_modules0 = ["to_q"]
+        target_modules1 = ["to_q", "to_k"]
+        with pytest.raises(RuntimeError):  # peft raises RuntimeError
+            with caplog.at_level(logging.ERROR):
+                self._check_model_hotswap(
+                    tmp_path,
+                    do_compile=True,
+                    rank0=8,
+                    rank1=8,
+                    target_modules0=target_modules0,
+                    target_modules1=target_modules1,
+                )
+                assert any("Hotswapping adapter0 was unsuccessful" in record.message for record in caplog.records)
+
+    @pytest.mark.parametrize("rank0,rank1", [(11, 11), (7, 13), (13, 7)])
+    @require_torch_version_greater("2.7.1")
+    def test_hotswapping_compile_on_different_shapes(self, tmp_path, rank0, rank1):
+        different_shapes_for_compilation = self.different_shapes_for_compilation
+        if different_shapes_for_compilation is None:
+            pytest.skip(f"Skipping as `different_shapes_for_compilation` is not set for {self.__class__.__name__}.")
+        # Specifying `use_duck_shape=False` instructs the compiler if it should use the same symbolic
+        # variable to represent input sizes that are the same. For more details,
+        # check out this [comment](https://github.com/huggingface/diffusers/pull/11327#discussion_r2047659790).
+        torch.fx.experimental._config.use_duck_shape = False
+
+        target_modules = ["to_q", "to_k", "to_v", "to_out.0"]
+        with torch._dynamo.config.patch(error_on_recompile=True):
+            self._check_model_hotswap(
+                tmp_path,
+                do_compile=True,
+                rank0=rank0,
+                rank1=rank1,
+                target_modules0=target_modules,
+            )

tests/models/testing_utils/memory.py

@@ -0,0 +1,498 @@
+# coding=utf-8
+# Copyright 2025 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 glob
+import inspect
+from functools import wraps
+
+import pytest
+import torch
+from accelerate.utils.modeling import compute_module_sizes
+
+from diffusers.utils.testing_utils import _check_safetensors_serialization
+from diffusers.utils.torch_utils import get_torch_cuda_device_capability
+
+from ...testing_utils import (
+    assert_tensors_close,
+    backend_empty_cache,
+    backend_max_memory_allocated,
+    backend_reset_peak_memory_stats,
+    backend_synchronize,
+    is_cpu_offload,
+    is_group_offload,
+    is_memory,
+    require_accelerator,
+    torch_device,
+)
+from .common import cast_inputs_to_dtype, check_device_map_is_respected
+
+
+def require_offload_support(func):
+    """
+    Decorator to skip tests if model doesn't support offloading (requires _no_split_modules).
+    """
+
+    @wraps(func)
+    def wrapper(self, *args, **kwargs):
+        if self.model_class._no_split_modules is None:
+            pytest.skip("Test not supported for this model as `_no_split_modules` is not set.")
+        return func(self, *args, **kwargs)
+
+    return wrapper
+
+
+def require_group_offload_support(func):
+    """
+    Decorator to skip tests if model doesn't support group offloading.
+    """
+
+    @wraps(func)
+    def wrapper(self, *args, **kwargs):
+        if not self.model_class._supports_group_offloading:
+            pytest.skip("Model does not support group offloading.")
+        return func(self, *args, **kwargs)
+
+    return wrapper
+
+
+@is_cpu_offload
+class CPUOffloadTesterMixin:
+    """
+    Mixin class for testing CPU offloading functionality.
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Optional properties:
+        - model_split_percents: List of percentages for splitting model across devices (default: [0.5, 0.7])
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: cpu_offload
+        Use `pytest -m "not cpu_offload"` to skip these tests
+    """
+
+    @property
+    def model_split_percents(self) -> list[float]:
+        """List of percentages for splitting model across devices during offloading tests."""
+        return [0.5, 0.7]
+
+    @require_offload_support
+    @torch.no_grad()
+    def test_cpu_offload(self, tmp_path, atol=1e-5, rtol=0):
+        config = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**config).eval()
+
+        model = model.to(torch_device)
+
+        torch.manual_seed(0)
+        base_output = model(**inputs_dict)
+
+        model_size = compute_module_sizes(model)[""]
+        # We test several splits of sizes to make sure it works
+        max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents]
+        model.cpu().save_pretrained(str(tmp_path))
+
+        for max_size in max_gpu_sizes:
+            max_memory = {0: max_size, "cpu": model_size * 2}
+            new_model = self.model_class.from_pretrained(str(tmp_path), device_map="auto", max_memory=max_memory)
+            # Making sure part of the model will actually end up offloaded
+            assert set(new_model.hf_device_map.values()) == {0, "cpu"}, "Model should be split between GPU and CPU"
+
+            check_device_map_is_respected(new_model, new_model.hf_device_map)
+            torch.manual_seed(0)
+            new_output = new_model(**inputs_dict)
+
+            assert_tensors_close(
+                base_output[0], new_output[0], atol=atol, rtol=rtol, msg="Output should match with CPU offloading"
+            )
+
+    @require_offload_support
+    @torch.no_grad()
+    def test_disk_offload_without_safetensors(self, tmp_path, atol=1e-5, rtol=0):
+        config = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**config).eval()
+
+        model = model.to(torch_device)
+
+        torch.manual_seed(0)
+        base_output = model(**inputs_dict)
+
+        model_size = compute_module_sizes(model)[""]
+        max_size = int(self.model_split_percents[0] * model_size)
+        # Force disk offload by setting very small CPU memory
+        max_memory = {0: max_size, "cpu": int(0.1 * max_size)}
+
+        model.cpu().save_pretrained(str(tmp_path), safe_serialization=False)
+        # This errors out because it's missing an offload folder
+        with pytest.raises(ValueError):
+            new_model = self.model_class.from_pretrained(str(tmp_path), device_map="auto", max_memory=max_memory)
+
+        new_model = self.model_class.from_pretrained(
+            str(tmp_path), device_map="auto", max_memory=max_memory, offload_folder=str(tmp_path)
+        )
+
+        check_device_map_is_respected(new_model, new_model.hf_device_map)
+        torch.manual_seed(0)
+        new_output = new_model(**inputs_dict)
+
+        assert_tensors_close(
+            base_output[0], new_output[0], atol=atol, rtol=rtol, msg="Output should match with disk offloading"
+        )
+
+    @require_offload_support
+    @torch.no_grad()
+    def test_disk_offload_with_safetensors(self, tmp_path, atol=1e-5, rtol=0):
+        config = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**config).eval()
+
+        model = model.to(torch_device)
+
+        torch.manual_seed(0)
+        base_output = model(**inputs_dict)
+
+        model_size = compute_module_sizes(model)[""]
+        model.cpu().save_pretrained(str(tmp_path))
+
+        max_size = int(self.model_split_percents[0] * model_size)
+        max_memory = {0: max_size, "cpu": max_size}
+        new_model = self.model_class.from_pretrained(
+            str(tmp_path), device_map="auto", offload_folder=str(tmp_path), max_memory=max_memory
+        )
+
+        check_device_map_is_respected(new_model, new_model.hf_device_map)
+        torch.manual_seed(0)
+        new_output = new_model(**inputs_dict)
+
+        assert_tensors_close(
+            base_output[0],
+            new_output[0],
+            atol=atol,
+            rtol=rtol,
+            msg="Output should match with disk offloading (safetensors)",
+        )
+
+
+@is_group_offload
+class GroupOffloadTesterMixin:
+    """
+    Mixin class for testing group offloading functionality.
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: group_offload
+        Use `pytest -m "not group_offload"` to skip these tests
+    """
+
+    @require_group_offload_support
+    @pytest.mark.parametrize("record_stream", [False, True])
+    def test_group_offloading(self, record_stream, atol=1e-5, rtol=0):
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        torch.manual_seed(0)
+
+        @torch.no_grad()
+        def run_forward(model):
+            assert all(
+                module._diffusers_hook.get_hook("group_offloading") is not None
+                for module in model.modules()
+                if hasattr(module, "_diffusers_hook")
+            ), "Group offloading hook should be set"
+            model.eval()
+            return model(**inputs_dict)[0]
+
+        model = self.model_class(**init_dict)
+
+        model.to(torch_device)
+        output_without_group_offloading = run_forward(model)
+
+        torch.manual_seed(0)
+        model = self.model_class(**init_dict)
+        model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1)
+        output_with_group_offloading1 = run_forward(model)
+
+        torch.manual_seed(0)
+        model = self.model_class(**init_dict)
+        model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1, non_blocking=True)
+        output_with_group_offloading2 = run_forward(model)
+
+        torch.manual_seed(0)
+        model = self.model_class(**init_dict)
+        model.enable_group_offload(torch_device, offload_type="leaf_level")
+        output_with_group_offloading3 = run_forward(model)
+
+        torch.manual_seed(0)
+        model = self.model_class(**init_dict)
+        model.enable_group_offload(
+            torch_device, offload_type="leaf_level", use_stream=True, record_stream=record_stream
+        )
+        output_with_group_offloading4 = run_forward(model)
+
+        assert_tensors_close(
+            output_without_group_offloading,
+            output_with_group_offloading1,
+            atol=atol,
+            rtol=rtol,
+            msg="Output should match with block-level offloading",
+        )
+        assert_tensors_close(
+            output_without_group_offloading,
+            output_with_group_offloading2,
+            atol=atol,
+            rtol=rtol,
+            msg="Output should match with non-blocking block-level offloading",
+        )
+        assert_tensors_close(
+            output_without_group_offloading,
+            output_with_group_offloading3,
+            atol=atol,
+            rtol=rtol,
+            msg="Output should match with leaf-level offloading",
+        )
+        assert_tensors_close(
+            output_without_group_offloading,
+            output_with_group_offloading4,
+            atol=atol,
+            rtol=rtol,
+            msg="Output should match with leaf-level offloading with stream",
+        )
+
+    @require_group_offload_support
+    @pytest.mark.parametrize("record_stream", [False, True])
+    @pytest.mark.parametrize("offload_type", ["block_level", "leaf_level"])
+    @torch.no_grad()
+    def test_group_offloading_with_layerwise_casting(self, record_stream, offload_type):
+        torch.manual_seed(0)
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        model = self.model_class(**init_dict)
+
+        model.to(torch_device)
+        model.eval()
+        _ = model(**inputs_dict)[0]
+
+        torch.manual_seed(0)
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        storage_dtype, compute_dtype = torch.float16, torch.float32
+        inputs_dict = cast_inputs_to_dtype(inputs_dict, torch.float32, compute_dtype)
+        model = self.model_class(**init_dict)
+        model.eval()
+        additional_kwargs = {} if offload_type == "leaf_level" else {"num_blocks_per_group": 1}
+        model.enable_group_offload(
+            torch_device, offload_type=offload_type, use_stream=True, record_stream=record_stream, **additional_kwargs
+        )
+        model.enable_layerwise_casting(storage_dtype=storage_dtype, compute_dtype=compute_dtype)
+        _ = model(**inputs_dict)[0]
+
+    @require_group_offload_support
+    @pytest.mark.parametrize("record_stream", [False, True])
+    @pytest.mark.parametrize("offload_type", ["block_level", "leaf_level"])
+    @torch.no_grad()
+    @torch.inference_mode()
+    def test_group_offloading_with_disk(self, tmp_path, record_stream, offload_type, atol=1e-5, rtol=0):
+        def _has_generator_arg(model):
+            sig = inspect.signature(model.forward)
+            params = sig.parameters
+            return "generator" in params
+
+        def _run_forward(model, inputs_dict):
+            accepts_generator = _has_generator_arg(model)
+            if accepts_generator:
+                inputs_dict["generator"] = torch.manual_seed(0)
+            torch.manual_seed(0)
+            return model(**inputs_dict)[0]
+
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        torch.manual_seed(0)
+        model = self.model_class(**init_dict)
+
+        model.eval()
+        model.to(torch_device)
+        output_without_group_offloading = _run_forward(model, inputs_dict)
+
+        torch.manual_seed(0)
+        model = self.model_class(**init_dict)
+        model.eval()
+
+        num_blocks_per_group = None if offload_type == "leaf_level" else 1
+        additional_kwargs = {} if offload_type == "leaf_level" else {"num_blocks_per_group": num_blocks_per_group}
+        tmpdir = str(tmp_path)
+        model.enable_group_offload(
+            torch_device,
+            offload_type=offload_type,
+            offload_to_disk_path=tmpdir,
+            use_stream=True,
+            record_stream=record_stream,
+            **additional_kwargs,
+        )
+        has_safetensors = glob.glob(f"{tmpdir}/*.safetensors")
+        assert has_safetensors, "No safetensors found in the directory."
+
+        # For "leaf-level", there is a prefetching hook which makes this check a bit non-deterministic
+        # in nature. So, skip it.
+        if offload_type != "leaf_level":
+            is_correct, extra_files, missing_files = _check_safetensors_serialization(
+                module=model,
+                offload_to_disk_path=tmpdir,
+                offload_type=offload_type,
+                num_blocks_per_group=num_blocks_per_group,
+            )
+            if not is_correct:
+                if extra_files:
+                    raise ValueError(f"Found extra files: {', '.join(extra_files)}")
+                elif missing_files:
+                    raise ValueError(f"Following files are missing: {', '.join(missing_files)}")
+
+        output_with_group_offloading = _run_forward(model, inputs_dict)
+        assert_tensors_close(
+            output_without_group_offloading,
+            output_with_group_offloading,
+            atol=atol,
+            rtol=rtol,
+            msg="Output should match with disk-based group offloading",
+        )
+
+
+class LayerwiseCastingTesterMixin:
+    """
+    Mixin class for testing layerwise dtype casting for memory optimization.
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+    """
+
+    @torch.no_grad()
+    def test_layerwise_casting_memory(self):
+        MB_TOLERANCE = 0.2
+        LEAST_COMPUTE_CAPABILITY = 8.0
+
+        def reset_memory_stats():
+            gc.collect()
+            backend_synchronize(torch_device)
+            backend_empty_cache(torch_device)
+            backend_reset_peak_memory_stats(torch_device)
+
+        def get_memory_usage(storage_dtype, compute_dtype):
+            torch.manual_seed(0)
+            config = self.get_init_dict()
+            inputs_dict = self.get_dummy_inputs()
+            inputs_dict = cast_inputs_to_dtype(inputs_dict, torch.float32, compute_dtype)
+            model = self.model_class(**config).eval()
+            model = model.to(torch_device, dtype=compute_dtype)
+            model.enable_layerwise_casting(storage_dtype=storage_dtype, compute_dtype=compute_dtype)
+
+            reset_memory_stats()
+            model(**inputs_dict)
+            model_memory_footprint = model.get_memory_footprint()
+            peak_inference_memory_allocated_mb = backend_max_memory_allocated(torch_device) / 1024**2
+
+            return model_memory_footprint, peak_inference_memory_allocated_mb
+
+        fp32_memory_footprint, fp32_max_memory = get_memory_usage(torch.float32, torch.float32)
+        fp8_e4m3_fp32_memory_footprint, fp8_e4m3_fp32_max_memory = get_memory_usage(torch.float8_e4m3fn, torch.float32)
+        fp8_e4m3_bf16_memory_footprint, fp8_e4m3_bf16_max_memory = get_memory_usage(
+            torch.float8_e4m3fn, torch.bfloat16
+        )
+
+        compute_capability = get_torch_cuda_device_capability() if torch_device == "cuda" else None
+        assert fp8_e4m3_bf16_memory_footprint < fp8_e4m3_fp32_memory_footprint < fp32_memory_footprint, (
+            "Memory footprint should decrease with lower precision storage"
+        )
+
+        # NOTE: the following assertion would fail on our CI (running Tesla T4) due to bf16 using more memory than fp32.
+        # On other devices, such as DGX (Ampere) and Audace (Ada), the test passes. So, we conditionally check it.
+        if compute_capability and compute_capability >= LEAST_COMPUTE_CAPABILITY:
+            assert fp8_e4m3_bf16_max_memory < fp8_e4m3_fp32_max_memory, (
+                "Peak memory should be lower with bf16 compute on newer GPUs"
+            )
+
+        # On this dummy test case with a small model, sometimes fp8_e4m3_fp32 max memory usage is higher than fp32 by a few
+        # bytes. This only happens for some models, so we allow a small tolerance.
+        # For any real model being tested, the order would be fp8_e4m3_bf16 < fp8_e4m3_fp32 < fp32.
+        assert (
+            fp8_e4m3_fp32_max_memory < fp32_max_memory
+            or abs(fp8_e4m3_fp32_max_memory - fp32_max_memory) < MB_TOLERANCE
+        ), "Peak memory should be lower or within tolerance with fp8 storage"
+
+    def test_layerwise_casting_training(self):
+        def test_fn(storage_dtype, compute_dtype):
+            if torch.device(torch_device).type == "cpu" and compute_dtype == torch.bfloat16:
+                pytest.skip("Skipping test because CPU doesn't go well with bfloat16.")
+
+            model = self.model_class(**self.get_init_dict())
+            model = model.to(torch_device, dtype=compute_dtype)
+            model.enable_layerwise_casting(storage_dtype=storage_dtype, compute_dtype=compute_dtype)
+            model.train()
+
+            inputs_dict = self.get_dummy_inputs()
+            inputs_dict = cast_inputs_to_dtype(inputs_dict, torch.float32, compute_dtype)
+            with torch.amp.autocast(device_type=torch.device(torch_device).type):
+                output = model(**inputs_dict, return_dict=False)[0]
+
+                input_tensor = inputs_dict[self.main_input_name]
+                noise = torch.randn((input_tensor.shape[0],) + self.output_shape).to(torch_device)
+                noise = cast_inputs_to_dtype(noise, torch.float32, compute_dtype)
+                loss = torch.nn.functional.mse_loss(output, noise)
+
+            loss.backward()
+
+        test_fn(torch.float16, torch.float32)
+        test_fn(torch.float8_e4m3fn, torch.float32)
+        test_fn(torch.float8_e5m2, torch.float32)
+        test_fn(torch.float8_e4m3fn, torch.bfloat16)
+
+
+@is_memory
+@require_accelerator
+class MemoryTesterMixin(CPUOffloadTesterMixin, GroupOffloadTesterMixin, LayerwiseCastingTesterMixin):
+    """
+    Combined mixin class for all memory optimization tests including CPU/disk offloading,
+    group offloading, and layerwise dtype casting.
+
+    This mixin inherits from:
+        - CPUOffloadTesterMixin: CPU and disk offloading tests
+        - GroupOffloadTesterMixin: Group offloading tests (block-level and leaf-level)
+        - LayerwiseCastingTesterMixin: Layerwise dtype casting tests
+
+    Expected from config mixin:
+        - model_class: The model class to test
+
+    Optional properties:
+        - model_split_percents: List of percentages for splitting model across devices (default: [0.5, 0.7])
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: memory
+        Use `pytest -m "not memory"` to skip these tests
+    """

tests/models/testing_utils/parallelism.py

@@ -0,0 +1,99 @@
+# coding=utf-8
+# Copyright 2025 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 os
+
+import pytest
+import torch
+import torch.multiprocessing as mp
+
+from diffusers.models._modeling_parallel import ContextParallelConfig
+
+from ...testing_utils import (
+    is_context_parallel,
+    require_torch_multi_accelerator,
+)
+
+
+@torch.no_grad()
+def _context_parallel_worker(rank, world_size, model_class, init_dict, cp_dict, inputs_dict, result_queue):
+    try:
+        os.environ["MASTER_ADDR"] = "localhost"
+        os.environ["MASTER_PORT"] = "12355"
+
+        torch.distributed.init_process_group(
+            backend="nccl",
+            init_method="env://",
+            world_size=world_size,
+            rank=rank,
+        )
+        torch.cuda.set_device(rank)
+        device = torch.device(f"cuda:{rank}")
+
+        model = model_class(**init_dict)
+        model.to(device)
+        model.eval()
+
+        inputs_on_device = {}
+        for key, value in inputs_dict.items():
+            if isinstance(value, torch.Tensor):
+                inputs_on_device[key] = value.to(device)
+            else:
+                inputs_on_device[key] = value
+
+        cp_config = ContextParallelConfig(**cp_dict)
+        model.enable_parallelism(config=cp_config)
+
+        output = model(**inputs_on_device, return_dict=False)[0]
+
+        if rank == 0:
+            result_queue.put(("success", output.shape))
+
+    except Exception as e:
+        if rank == 0:
+            result_queue.put(("error", str(e)))
+    finally:
+        if torch.distributed.is_initialized():
+            torch.distributed.destroy_process_group()
+
+
+@is_context_parallel
+@require_torch_multi_accelerator
+class ContextParallelTesterMixin:
+    @pytest.mark.parametrize("cp_type", ["ulysses_degree", "ring_degree"], ids=["ulysses", "ring"])
+    def test_context_parallel_inference(self, cp_type):
+        if not torch.distributed.is_available():
+            pytest.skip("torch.distributed is not available.")
+
+        if not hasattr(self.model_class, "_cp_plan") or self.model_class._cp_plan is None:
+            pytest.skip("Model does not have a _cp_plan defined for context parallel inference.")
+
+        world_size = 2
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        cp_dict = {cp_type: world_size}
+
+        ctx = mp.get_context("spawn")
+        result_queue = ctx.Queue()
+
+        mp.spawn(
+            _context_parallel_worker,
+            args=(world_size, self.model_class, init_dict, cp_dict, inputs_dict, result_queue),
+            nprocs=world_size,
+            join=True,
+        )
+
+        status, result = result_queue.get(timeout=60)
+        assert status == "success", f"Context parallel inference failed: {result}"

tests/models/testing_utils/single_file.py

@@ -0,0 +1,265 @@
+# coding=utf-8
+# Copyright 2025 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 torch
+from huggingface_hub import hf_hub_download, snapshot_download
+
+from diffusers.loaders.single_file_utils import _extract_repo_id_and_weights_name
+
+from ...testing_utils import (
+    assert_tensors_close,
+    backend_empty_cache,
+    is_single_file,
+    nightly,
+    require_torch_accelerator,
+    torch_device,
+)
+from .common import check_device_map_is_respected
+
+
+def download_single_file_checkpoint(pretrained_model_name_or_path, filename, tmpdir):
+    """Download a single file checkpoint from the Hub to a temporary directory."""
+    path = hf_hub_download(pretrained_model_name_or_path, filename=filename, local_dir=tmpdir)
+    return path
+
+
+def download_diffusers_config(pretrained_model_name_or_path, tmpdir):
+    """Download diffusers config files (excluding weights) from a repository."""
+    path = snapshot_download(
+        pretrained_model_name_or_path,
+        ignore_patterns=[
+            "**/*.ckpt",
+            "*.ckpt",
+            "**/*.bin",
+            "*.bin",
+            "**/*.pt",
+            "*.pt",
+            "**/*.safetensors",
+            "*.safetensors",
+        ],
+        allow_patterns=["**/*.json", "*.json", "*.txt", "**/*.txt"],
+        local_dir=tmpdir,
+    )
+    return path
+
+
+@nightly
+@require_torch_accelerator
+@is_single_file
+class SingleFileTesterMixin:
+    """
+    Mixin class for testing single file loading for models.
+
+    Required properties (must be implemented by subclasses):
+        - ckpt_path: Path or Hub path to the single file checkpoint
+
+    Optional properties:
+        - torch_dtype: torch dtype to use for testing (default: None)
+        - alternate_ckpt_paths: List of alternate checkpoint paths for variant testing (default: None)
+
+    Expected from config mixin:
+        - model_class: The model class to test
+        - pretrained_model_name_or_path: Hub repository ID for the pretrained model
+        - pretrained_model_kwargs: Additional kwargs for from_pretrained (e.g., subfolder)
+
+    Pytest mark: single_file
+        Use `pytest -m "not single_file"` to skip these tests
+    """
+
+    # ==================== Required Properties ====================
+
+    @property
+    def ckpt_path(self) -> str:
+        """Path or Hub path to the single file checkpoint. Must be implemented by subclasses."""
+        raise NotImplementedError("Subclasses must implement the `ckpt_path` property.")
+
+    # ==================== Optional Properties ====================
+
+    @property
+    def torch_dtype(self) -> torch.dtype | None:
+        """torch dtype to use for single file testing."""
+        return None
+
+    @property
+    def alternate_ckpt_paths(self) -> list[str] | None:
+        """List of alternate checkpoint paths for variant testing."""
+        return None
+
+    def setup_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def teardown_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def test_single_file_model_config(self):
+        pretrained_kwargs = {"device": torch_device, **self.pretrained_model_kwargs}
+        single_file_kwargs = {"device": torch_device}
+
+        if self.torch_dtype:
+            pretrained_kwargs["torch_dtype"] = self.torch_dtype
+            single_file_kwargs["torch_dtype"] = self.torch_dtype
+
+        model = self.model_class.from_pretrained(self.pretrained_model_name_or_path, **pretrained_kwargs)
+        model_single_file = self.model_class.from_single_file(self.ckpt_path, **single_file_kwargs)
+
+        PARAMS_TO_IGNORE = ["torch_dtype", "_name_or_path", "_use_default_values", "_diffusers_version"]
+        for param_name, param_value in model_single_file.config.items():
+            if param_name in PARAMS_TO_IGNORE:
+                continue
+            assert model.config[param_name] == param_value, (
+                f"{param_name} differs between pretrained loading and single file loading: "
+                f"pretrained={model.config[param_name]}, single_file={param_value}"
+            )
+
+    def test_single_file_model_parameters(self, atol=1e-5, rtol=1e-5):
+        pretrained_kwargs = {"device": torch_device, **self.pretrained_model_kwargs}
+        single_file_kwargs = {"device": torch_device}
+
+        if self.torch_dtype:
+            pretrained_kwargs["torch_dtype"] = self.torch_dtype
+            single_file_kwargs["torch_dtype"] = self.torch_dtype
+
+        model = self.model_class.from_pretrained(self.pretrained_model_name_or_path, **pretrained_kwargs)
+        model_single_file = self.model_class.from_single_file(self.ckpt_path, **single_file_kwargs)
+
+        state_dict = model.state_dict()
+        state_dict_single_file = model_single_file.state_dict()
+
+        assert set(state_dict.keys()) == set(state_dict_single_file.keys()), (
+            "Model parameters keys differ between pretrained and single file loading. "
+            f"Missing in single file: {set(state_dict.keys()) - set(state_dict_single_file.keys())}. "
+            f"Extra in single file: {set(state_dict_single_file.keys()) - set(state_dict.keys())}"
+        )
+
+        for key in state_dict.keys():
+            param = state_dict[key]
+            param_single_file = state_dict_single_file[key]
+
+            assert param.shape == param_single_file.shape, (
+                f"Parameter shape mismatch for {key}: "
+                f"pretrained {param.shape} vs single file {param_single_file.shape}"
+            )
+
+            assert_tensors_close(
+                param, param_single_file, atol=atol, rtol=rtol, msg=f"Parameter values differ for {key}"
+            )
+
+    def test_single_file_loading_local_files_only(self, tmp_path):
+        single_file_kwargs = {}
+
+        if self.torch_dtype:
+            single_file_kwargs["torch_dtype"] = self.torch_dtype
+
+        pretrained_model_name_or_path, weight_name = _extract_repo_id_and_weights_name(self.ckpt_path)
+        local_ckpt_path = download_single_file_checkpoint(pretrained_model_name_or_path, weight_name, str(tmp_path))
+
+        model_single_file = self.model_class.from_single_file(
+            local_ckpt_path, local_files_only=True, **single_file_kwargs
+        )
+
+        assert model_single_file is not None, "Failed to load model with local_files_only=True"
+
+    def test_single_file_loading_with_diffusers_config(self):
+        single_file_kwargs = {}
+
+        if self.torch_dtype:
+            single_file_kwargs["torch_dtype"] = self.torch_dtype
+
+        # Load with config parameter
+        model_single_file = self.model_class.from_single_file(
+            self.ckpt_path, config=self.pretrained_model_name_or_path, **single_file_kwargs
+        )
+
+        # Load pretrained for comparison
+        pretrained_kwargs = {**self.pretrained_model_kwargs}
+        if self.torch_dtype:
+            pretrained_kwargs["torch_dtype"] = self.torch_dtype
+
+        model = self.model_class.from_pretrained(self.pretrained_model_name_or_path, **pretrained_kwargs)
+
+        # Compare configs
+        PARAMS_TO_IGNORE = ["torch_dtype", "_name_or_path", "_use_default_values", "_diffusers_version"]
+        for param_name, param_value in model_single_file.config.items():
+            if param_name in PARAMS_TO_IGNORE:
+                continue
+            assert model.config[param_name] == param_value, (
+                f"{param_name} differs: pretrained={model.config[param_name]}, single_file={param_value}"
+            )
+
+    def test_single_file_loading_with_diffusers_config_local_files_only(self, tmp_path):
+        single_file_kwargs = {}
+
+        if self.torch_dtype:
+            single_file_kwargs["torch_dtype"] = self.torch_dtype
+
+        pretrained_model_name_or_path, weight_name = _extract_repo_id_and_weights_name(self.ckpt_path)
+        local_ckpt_path = download_single_file_checkpoint(pretrained_model_name_or_path, weight_name, str(tmp_path))
+        local_diffusers_config = download_diffusers_config(self.pretrained_model_name_or_path, str(tmp_path))
+
+        model_single_file = self.model_class.from_single_file(
+            local_ckpt_path, config=local_diffusers_config, local_files_only=True, **single_file_kwargs
+        )
+
+        assert model_single_file is not None, "Failed to load model with config and local_files_only=True"
+
+    def test_single_file_loading_dtype(self):
+        for dtype in [torch.float32, torch.float16]:
+            if torch_device == "mps" and dtype == torch.bfloat16:
+                continue
+
+            model_single_file = self.model_class.from_single_file(self.ckpt_path, torch_dtype=dtype)
+
+            assert model_single_file.dtype == dtype, f"Expected dtype {dtype}, got {model_single_file.dtype}"
+
+            # Cleanup
+            del model_single_file
+            gc.collect()
+            backend_empty_cache(torch_device)
+
+    def test_checkpoint_variant_loading(self):
+        if not self.alternate_ckpt_paths:
+            return
+
+        for ckpt_path in self.alternate_ckpt_paths:
+            backend_empty_cache(torch_device)
+
+            single_file_kwargs = {}
+            if self.torch_dtype:
+                single_file_kwargs["torch_dtype"] = self.torch_dtype
+
+            model = self.model_class.from_single_file(ckpt_path, **single_file_kwargs)
+
+            assert model is not None, f"Failed to load checkpoint from {ckpt_path}"
+
+            del model
+            gc.collect()
+            backend_empty_cache(torch_device)
+
+    def test_single_file_loading_with_device_map(self):
+        single_file_kwargs = {"device_map": torch_device}
+
+        if self.torch_dtype:
+            single_file_kwargs["torch_dtype"] = self.torch_dtype
+
+        model = self.model_class.from_single_file(self.ckpt_path, **single_file_kwargs)
+
+        assert model is not None, "Failed to load model with device_map"
+        assert hasattr(model, "hf_device_map"), "Model should have hf_device_map attribute when loaded with device_map"
+        assert model.hf_device_map is not None, "hf_device_map should not be None when loaded with device_map"
+        check_device_map_is_respected(model, model.hf_device_map)

tests/models/testing_utils/training.py

@@ -0,0 +1,220 @@
+# coding=utf-8
+# Copyright 2025 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 copy
+import gc
+
+import pytest
+import torch
+
+from diffusers.training_utils import EMAModel
+
+from ...testing_utils import (
+    backend_empty_cache,
+    is_training,
+    require_torch_accelerator_with_training,
+    torch_all_close,
+    torch_device,
+)
+
+
+@is_training
+@require_torch_accelerator_with_training
+class TrainingTesterMixin:
+    """
+    Mixin class for testing training functionality on models.
+
+    Expected from config mixin:
+        - model_class: The model class to test
+        - output_shape: Tuple defining the expected output shape
+
+    Expected methods from config mixin:
+        - get_init_dict(): Returns dict of arguments to initialize the model
+        - get_dummy_inputs(): Returns dict of inputs to pass to the model forward pass
+
+    Pytest mark: training
+        Use `pytest -m "not training"` to skip these tests
+    """
+
+    def setup_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def teardown_method(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def test_training(self):
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.train()
+        output = model(**inputs_dict, return_dict=False)[0]
+
+        noise = torch.randn((output.shape[0],) + self.output_shape).to(torch_device)
+        loss = torch.nn.functional.mse_loss(output, noise)
+        loss.backward()
+
+    def test_training_with_ema(self):
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.train()
+        ema_model = EMAModel(model.parameters())
+
+        output = model(**inputs_dict, return_dict=False)[0]
+
+        noise = torch.randn((output.shape[0],) + self.output_shape).to(torch_device)
+        loss = torch.nn.functional.mse_loss(output, noise)
+        loss.backward()
+        ema_model.step(model.parameters())
+
+    def test_gradient_checkpointing(self):
+        if not self.model_class._supports_gradient_checkpointing:
+            pytest.skip("Gradient checkpointing is not supported.")
+
+        init_dict = self.get_init_dict()
+
+        # at init model should have gradient checkpointing disabled
+        model = self.model_class(**init_dict)
+        assert not model.is_gradient_checkpointing, "Gradient checkpointing should be disabled at init"
+
+        # check enable works
+        model.enable_gradient_checkpointing()
+        assert model.is_gradient_checkpointing, "Gradient checkpointing should be enabled"
+
+        # check disable works
+        model.disable_gradient_checkpointing()
+        assert not model.is_gradient_checkpointing, "Gradient checkpointing should be disabled"
+
+    def test_gradient_checkpointing_is_applied(self, expected_set=None):
+        if not self.model_class._supports_gradient_checkpointing:
+            pytest.skip("Gradient checkpointing is not supported.")
+
+        if expected_set is None:
+            pytest.skip("expected_set must be provided to verify gradient checkpointing is applied.")
+
+        init_dict = self.get_init_dict()
+
+        model_class_copy = copy.copy(self.model_class)
+        model = model_class_copy(**init_dict)
+        model.enable_gradient_checkpointing()
+
+        modules_with_gc_enabled = {}
+        for submodule in model.modules():
+            if hasattr(submodule, "gradient_checkpointing"):
+                assert submodule.gradient_checkpointing, f"{submodule.__class__.__name__} should have GC enabled"
+                modules_with_gc_enabled[submodule.__class__.__name__] = True
+
+        assert set(modules_with_gc_enabled.keys()) == expected_set, (
+            f"Modules with GC enabled {set(modules_with_gc_enabled.keys())} do not match expected set {expected_set}"
+        )
+        assert all(modules_with_gc_enabled.values()), "All modules should have GC enabled"
+
+    def test_gradient_checkpointing_equivalence(self, loss_tolerance=1e-5, param_grad_tol=5e-5, skip=None):
+        if not self.model_class._supports_gradient_checkpointing:
+            pytest.skip("Gradient checkpointing is not supported.")
+
+        if skip is None:
+            skip = set()
+
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+        inputs_dict_copy = copy.deepcopy(inputs_dict)
+
+        torch.manual_seed(0)
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+
+        assert not model.is_gradient_checkpointing and model.training
+
+        out = model(**inputs_dict, return_dict=False)[0]
+
+        # run the backwards pass on the model
+        model.zero_grad()
+
+        labels = torch.randn_like(out)
+        loss = (out - labels).mean()
+        loss.backward()
+
+        # re-instantiate the model now enabling gradient checkpointing
+        torch.manual_seed(0)
+        model_2 = self.model_class(**init_dict)
+        # clone model
+        model_2.load_state_dict(model.state_dict())
+        model_2.to(torch_device)
+        model_2.enable_gradient_checkpointing()
+
+        assert model_2.is_gradient_checkpointing and model_2.training
+
+        out_2 = model_2(**inputs_dict_copy, return_dict=False)[0]
+
+        # run the backwards pass on the model
+        model_2.zero_grad()
+        loss_2 = (out_2 - labels).mean()
+        loss_2.backward()
+
+        # compare the output and parameters gradients
+        assert (loss - loss_2).abs() < loss_tolerance, (
+            f"Loss difference {(loss - loss_2).abs()} exceeds tolerance {loss_tolerance}"
+        )
+
+        named_params = dict(model.named_parameters())
+        named_params_2 = dict(model_2.named_parameters())
+
+        for name, param in named_params.items():
+            if "post_quant_conv" in name:
+                continue
+            if name in skip:
+                continue
+            if param.grad is None:
+                continue
+
+            assert torch_all_close(param.grad.data, named_params_2[name].grad.data, atol=param_grad_tol), (
+                f"Gradient mismatch for {name}"
+            )
+
+    def test_mixed_precision_training(self):
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.train()
+
+        # Test with float16
+        if torch.device(torch_device).type != "cpu":
+            with torch.amp.autocast(device_type=torch.device(torch_device).type, dtype=torch.float16):
+                output = model(**inputs_dict, return_dict=False)[0]
+
+                noise = torch.randn((output.shape[0],) + self.output_shape).to(torch_device)
+                loss = torch.nn.functional.mse_loss(output, noise)
+
+            loss.backward()
+
+        # Test with bfloat16
+        if torch.device(torch_device).type != "cpu":
+            model.zero_grad()
+            with torch.amp.autocast(device_type=torch.device(torch_device).type, dtype=torch.bfloat16):
+                output = model(**inputs_dict, return_dict=False)[0]
+
+                noise = torch.randn((output.shape[0],) + self.output_shape).to(torch_device)
+                loss = torch.nn.functional.mse_loss(output, noise)
+
+            loss.backward()

tests/models/transformers/test_models_transformer_flux.py

@@ -13,23 +13,51 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import unittest
+from typing import Any
 
 import torch
 
 from diffusers import FluxTransformer2DModel
-from diffusers.models.attention_processor import FluxIPAdapterJointAttnProcessor2_0
 from diffusers.models.embeddings import ImageProjection
+from diffusers.models.transformers.transformer_flux import FluxIPAdapterAttnProcessor
+from diffusers.utils.torch_utils import randn_tensor
 
-from ...testing_utils import enable_full_determinism, is_peft_available, torch_device
-from ..test_modeling_common import LoraHotSwappingForModelTesterMixin, ModelTesterMixin, TorchCompileTesterMixin
+from ...testing_utils import enable_full_determinism, torch_device
+from ..testing_utils import (
+    AttentionTesterMixin,
+    BaseModelTesterConfig,
+    BitsAndBytesCompileTesterMixin,
+    BitsAndBytesTesterMixin,
+    ContextParallelTesterMixin,
+    FasterCacheTesterMixin,
+    FirstBlockCacheTesterMixin,
+    GGUFCompileTesterMixin,
+    GGUFTesterMixin,
+    IPAdapterTesterMixin,
+    LoraHotSwappingForModelTesterMixin,
+    LoraTesterMixin,
+    MemoryTesterMixin,
+    ModelOptCompileTesterMixin,
+    ModelOptTesterMixin,
+    ModelTesterMixin,
+    PyramidAttentionBroadcastTesterMixin,
+    QuantoCompileTesterMixin,
+    QuantoTesterMixin,
+    SingleFileTesterMixin,
+    TorchAoCompileTesterMixin,
+    TorchAoTesterMixin,
+    TorchCompileTesterMixin,
+    TrainingTesterMixin,
+)
 
 
 enable_full_determinism()
 
 
-def create_flux_ip_adapter_state_dict(model):
-    # "ip_adapter" (cross-attention weights)
+# TODO: This standalone function maintains backward compatibility with pipeline tests
+# (tests/pipelines/test_pipelines_common.py) and will be refactored.
+def create_flux_ip_adapter_state_dict(model) -> dict[str, dict[str, Any]]:
+    """Create a dummy IP Adapter state dict for Flux transformer testing."""
     ip_cross_attn_state_dict = {}
     key_id = 0
 
@@ -39,7 +67,7 @@ def create_flux_ip_adapter_state_dict(model):
 
         joint_attention_dim = model.config["joint_attention_dim"]
         hidden_size = model.config["num_attention_heads"] * model.config["attention_head_dim"]
-        sd = FluxIPAdapterJointAttnProcessor2_0(
+        sd = FluxIPAdapterAttnProcessor(
             hidden_size=hidden_size, cross_attention_dim=joint_attention_dim, scale=1.0
         ).state_dict()
         ip_cross_attn_state_dict.update(
@@ -50,11 +78,8 @@ def create_flux_ip_adapter_state_dict(model):
                 f"{key_id}.to_v_ip.bias": sd["to_v_ip.0.bias"],
             }
         )
-
         key_id += 1
 
-    # "image_proj" (ImageProjection layer weights)
-
     image_projection = ImageProjection(
         cross_attention_dim=model.config["joint_attention_dim"],
         image_embed_dim=(
@@ -75,57 +100,37 @@ def create_flux_ip_adapter_state_dict(model):
     )
 
     del sd
-    ip_state_dict = {}
-    ip_state_dict.update({"image_proj": ip_image_projection_state_dict, "ip_adapter": ip_cross_attn_state_dict})
-    return ip_state_dict
+    return {"image_proj": ip_image_projection_state_dict, "ip_adapter": ip_cross_attn_state_dict}
 
 
-class FluxTransformerTests(ModelTesterMixin, unittest.TestCase):
-    model_class = FluxTransformer2DModel
-    main_input_name = "hidden_states"
-    # We override the items here because the transformer under consideration is small.
-    model_split_percents = [0.7, 0.6, 0.6]
-
-    # Skip setting testing with default: AttnProcessor
-    uses_custom_attn_processor = True
+class FluxTransformerTesterConfig(BaseModelTesterConfig):
+    @property
+    def model_class(self):
+        return FluxTransformer2DModel
 
     @property
-    def dummy_input(self):
-        return self.prepare_dummy_input()
+    def pretrained_model_name_or_path(self):
+        return "hf-internal-testing/tiny-flux-pipe"
 
     @property
-    def input_shape(self):
+    def pretrained_model_kwargs(self):
+        return {"subfolder": "transformer"}
+
+    @property
+    def output_shape(self) -> tuple[int, int]:
         return (16, 4)
 
     @property
-    def output_shape(self):
+    def input_shape(self) -> tuple[int, int]:
         return (16, 4)
 
-    def prepare_dummy_input(self, height=4, width=4):
-        batch_size = 1
-        num_latent_channels = 4
-        num_image_channels = 3
-        sequence_length = 48
-        embedding_dim = 32
-
-        hidden_states = torch.randn((batch_size, height * width, num_latent_channels)).to(torch_device)
-        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
-        pooled_prompt_embeds = torch.randn((batch_size, embedding_dim)).to(torch_device)
-        text_ids = torch.randn((sequence_length, num_image_channels)).to(torch_device)
-        image_ids = torch.randn((height * width, num_image_channels)).to(torch_device)
-        timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)
+    @property
+    def generator(self):
+        return torch.Generator("cpu").manual_seed(0)
 
+    def get_init_dict(self) -> dict[str, int | list[int]]:
+        """Return Flux model initialization arguments."""
         return {
-            "hidden_states": hidden_states,
-            "encoder_hidden_states": encoder_hidden_states,
-            "img_ids": image_ids,
-            "txt_ids": text_ids,
-            "pooled_projections": pooled_prompt_embeds,
-            "timestep": timestep,
-        }
-
-    def prepare_init_args_and_inputs_for_common(self):
-        init_dict = {
             "patch_size": 1,
             "in_channels": 4,
             "num_layers": 1,
@@ -137,11 +142,40 @@ class FluxTransformerTests(ModelTesterMixin, unittest.TestCase):
             "axes_dims_rope": [4, 4, 8],
         }
 
-        inputs_dict = self.dummy_input
-        return init_dict, inputs_dict
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        batch_size = 1
+        height = width = 4
+        num_latent_channels = 4
+        num_image_channels = 3
+        sequence_length = 48
+        embedding_dim = 32
 
+        return {
+            "hidden_states": randn_tensor(
+                (batch_size, height * width, num_latent_channels), generator=self.generator, device=torch_device
+            ),
+            "encoder_hidden_states": randn_tensor(
+                (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
+            ),
+            "pooled_projections": randn_tensor(
+                (batch_size, embedding_dim), generator=self.generator, device=torch_device
+            ),
+            "img_ids": randn_tensor(
+                (height * width, num_image_channels), generator=self.generator, device=torch_device
+            ),
+            "txt_ids": randn_tensor(
+                (sequence_length, num_image_channels), generator=self.generator, device=torch_device
+            ),
+            "timestep": torch.tensor([1.0]).to(torch_device).expand(batch_size),
+        }
+
+
+class TestFluxTransformer(FluxTransformerTesterConfig, ModelTesterMixin):
     def test_deprecated_inputs_img_txt_ids_3d(self):
-        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        """Test that deprecated 3D img_ids and txt_ids still work."""
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+
         model = self.model_class(**init_dict)
         model.to(torch_device)
         model.eval()
@@ -162,63 +196,269 @@ class FluxTransformerTests(ModelTesterMixin, unittest.TestCase):
         with torch.no_grad():
             output_2 = model(**inputs_dict).to_tuple()[0]
 
-        self.assertEqual(output_1.shape, output_2.shape)
-        self.assertTrue(
-            torch.allclose(output_1, output_2, atol=1e-5),
-            msg="output with deprecated inputs (img_ids and txt_ids as 3d torch tensors) are not equal as them as 2d inputs",
+        assert output_1.shape == output_2.shape
+        assert torch.allclose(output_1, output_2, atol=1e-5), (
+            "output with deprecated inputs (img_ids and txt_ids as 3d torch tensors) "
+            "are not equal as them as 2d inputs"
         )
 
-    def test_gradient_checkpointing_is_applied(self):
-        expected_set = {"FluxTransformer2DModel"}
-        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
-    # The test exists for cases like
-    # https://github.com/huggingface/diffusers/issues/11874
-    @unittest.skipIf(not is_peft_available(), "Only with PEFT")
-    def test_lora_exclude_modules(self):
-        from peft import LoraConfig, get_peft_model_state_dict, inject_adapter_in_model, set_peft_model_state_dict
+class TestFluxTransformerMemory(FluxTransformerTesterConfig, MemoryTesterMixin):
+    """Memory optimization tests for Flux Transformer."""
 
-        lora_rank = 4
-        target_module = "single_transformer_blocks.0.proj_out"
-        adapter_name = "foo"
-        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
-        model = self.model_class(**init_dict).to(torch_device)
 
-        state_dict = model.state_dict()
-        target_mod_shape = state_dict[f"{target_module}.weight"].shape
-        lora_state_dict = {
-            f"{target_module}.lora_A.weight": torch.ones(lora_rank, target_mod_shape[1]) * 22,
-            f"{target_module}.lora_B.weight": torch.ones(target_mod_shape[0], lora_rank) * 33,
+class TestFluxTransformerTraining(FluxTransformerTesterConfig, TrainingTesterMixin):
+    """Training tests for Flux Transformer."""
+
+
+class TestFluxTransformerAttention(FluxTransformerTesterConfig, AttentionTesterMixin):
+    """Attention processor tests for Flux Transformer."""
+
+
+class TestFluxTransformerContextParallel(FluxTransformerTesterConfig, ContextParallelTesterMixin):
+    """Context Parallel inference tests for Flux Transformer"""
+
+
+class TestFluxTransformerIPAdapter(FluxTransformerTesterConfig, IPAdapterTesterMixin):
+    """IP Adapter tests for Flux Transformer."""
+
+    @property
+    def ip_adapter_processor_cls(self):
+        return FluxIPAdapterAttnProcessor
+
+    def modify_inputs_for_ip_adapter(self, model, inputs_dict):
+        torch.manual_seed(0)
+        # Create dummy image embeds for IP adapter
+        cross_attention_dim = getattr(model.config, "joint_attention_dim", 32)
+        image_embeds = torch.randn(1, 1, cross_attention_dim).to(torch_device)
+
+        inputs_dict.update({"joint_attention_kwargs": {"ip_adapter_image_embeds": image_embeds}})
+
+        return inputs_dict
+
+    def create_ip_adapter_state_dict(self, model: Any) -> dict[str, dict[str, Any]]:
+        return create_flux_ip_adapter_state_dict(model)
+
+
+class TestFluxTransformerLoRA(FluxTransformerTesterConfig, LoraTesterMixin):
+    """LoRA adapter tests for Flux Transformer."""
+
+
+class TestFluxTransformerLoRAHotSwap(FluxTransformerTesterConfig, LoraHotSwappingForModelTesterMixin):
+    """LoRA hot-swapping tests for Flux Transformer."""
+
+    @property
+    def different_shapes_for_compilation(self):
+        return [(4, 4), (4, 8), (8, 8)]
+
+    def get_dummy_inputs(self, height: int = 4, width: int = 4) -> dict[str, torch.Tensor]:
+        """Override to support dynamic height/width for LoRA hotswap tests."""
+        batch_size = 1
+        num_latent_channels = 4
+        num_image_channels = 3
+        sequence_length = 24
+        embedding_dim = 8
+
+        return {
+            "hidden_states": randn_tensor((batch_size, height * width, num_latent_channels), device=torch_device),
+            "encoder_hidden_states": randn_tensor((batch_size, sequence_length, embedding_dim), device=torch_device),
+            "pooled_projections": randn_tensor((batch_size, embedding_dim), device=torch_device),
+            "img_ids": randn_tensor((height * width, num_image_channels), device=torch_device),
+            "txt_ids": randn_tensor((sequence_length, num_image_channels), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device).expand(batch_size),
         }
-        # Passing exclude_modules should no longer be necessary (or even passing target_modules, for that matter).
-        config = LoraConfig(
-            r=lora_rank, target_modules=["single_transformer_blocks.0.proj_out"], exclude_modules=["proj_out"]
-        )
-        inject_adapter_in_model(config, model, adapter_name=adapter_name, state_dict=lora_state_dict)
-        set_peft_model_state_dict(model, lora_state_dict, adapter_name)
-        retrieved_lora_state_dict = get_peft_model_state_dict(model, adapter_name=adapter_name)
-        assert len(retrieved_lora_state_dict) == len(lora_state_dict)
-        assert (retrieved_lora_state_dict["single_transformer_blocks.0.proj_out.lora_A.weight"] == 22).all()
-        assert (retrieved_lora_state_dict["single_transformer_blocks.0.proj_out.lora_B.weight"] == 33).all()
 
 
-class FluxTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
-    model_class = FluxTransformer2DModel
-    different_shapes_for_compilation = [(4, 4), (4, 8), (8, 8)]
+class TestFluxTransformerCompile(FluxTransformerTesterConfig, TorchCompileTesterMixin):
+    @property
+    def different_shapes_for_compilation(self):
+        return [(4, 4), (4, 8), (8, 8)]
 
-    def prepare_init_args_and_inputs_for_common(self):
-        return FluxTransformerTests().prepare_init_args_and_inputs_for_common()
+    def get_dummy_inputs(self, height: int = 4, width: int = 4) -> dict[str, torch.Tensor]:
+        """Override to support dynamic height/width for compilation tests."""
+        batch_size = 1
+        num_latent_channels = 4
+        num_image_channels = 3
+        sequence_length = 24
+        embedding_dim = 8
 
-    def prepare_dummy_input(self, height, width):
-        return FluxTransformerTests().prepare_dummy_input(height=height, width=width)
+        return {
+            "hidden_states": randn_tensor((batch_size, height * width, num_latent_channels), device=torch_device),
+            "encoder_hidden_states": randn_tensor((batch_size, sequence_length, embedding_dim), device=torch_device),
+            "pooled_projections": randn_tensor((batch_size, embedding_dim), device=torch_device),
+            "img_ids": randn_tensor((height * width, num_image_channels), device=torch_device),
+            "txt_ids": randn_tensor((sequence_length, num_image_channels), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device).expand(batch_size),
+        }
 
 
-class FluxTransformerLoRAHotSwapTests(LoraHotSwappingForModelTesterMixin, unittest.TestCase):
-    model_class = FluxTransformer2DModel
-    different_shapes_for_compilation = [(4, 4), (4, 8), (8, 8)]
+class TestFluxSingleFile(FluxTransformerTesterConfig, SingleFileTesterMixin):
+    @property
+    def ckpt_path(self):
+        return "https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors"
 
-    def prepare_init_args_and_inputs_for_common(self):
-        return FluxTransformerTests().prepare_init_args_and_inputs_for_common()
+    @property
+    def alternate_ckpt_paths(self):
+        return ["https://huggingface.co/Comfy-Org/flux1-dev/blob/main/flux1-dev-fp8.safetensors"]
 
-    def prepare_dummy_input(self, height, width):
-        return FluxTransformerTests().prepare_dummy_input(height=height, width=width)
+    @property
+    def pretrained_model_name_or_path(self):
+        return "black-forest-labs/FLUX.1-dev"
+
+
+class TestFluxTransformerBitsAndBytes(FluxTransformerTesterConfig, BitsAndBytesTesterMixin):
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerQuanto(FluxTransformerTesterConfig, QuantoTesterMixin):
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerTorchAo(FluxTransformerTesterConfig, TorchAoTesterMixin):
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerGGUF(FluxTransformerTesterConfig, GGUFTesterMixin):
+    @property
+    def gguf_filename(self):
+        return "https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q8_0.gguf"
+
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerModelOpt(FluxTransformerTesterConfig, ModelOptTesterMixin):
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerBitsAndBytesCompile(FluxTransformerTesterConfig, BitsAndBytesCompileTesterMixin):
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerQuantoCompile(FluxTransformerTesterConfig, QuantoCompileTesterMixin):
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerTorchAoCompile(FluxTransformerTesterConfig, TorchAoCompileTesterMixin):
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerGGUFCompile(FluxTransformerTesterConfig, GGUFCompileTesterMixin):
+    @property
+    def gguf_filename(self):
+        return "https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q8_0.gguf"
+
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerModelOptCompile(FluxTransformerTesterConfig, ModelOptCompileTesterMixin):
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        return {
+            "hidden_states": randn_tensor((1, 4096, 64), device=torch_device),
+            "encoder_hidden_states": randn_tensor((1, 512, 4096), device=torch_device),
+            "pooled_projections": randn_tensor((1, 768), device=torch_device),
+            "timestep": torch.tensor([1.0]).to(torch_device),
+            "img_ids": randn_tensor((4096, 3), device=torch_device),
+            "txt_ids": randn_tensor((512, 3), device=torch_device),
+            "guidance": torch.tensor([3.5]).to(torch_device),
+        }
+
+
+class TestFluxTransformerPABCache(FluxTransformerTesterConfig, PyramidAttentionBroadcastTesterMixin):
+    """PyramidAttentionBroadcast cache tests for Flux Transformer."""
+
+
+class TestFluxTransformerFBCCache(FluxTransformerTesterConfig, FirstBlockCacheTesterMixin):
+    """FirstBlockCache tests for Flux Transformer."""
+
+
+class TestFluxTransformerFasterCache(FluxTransformerTesterConfig, FasterCacheTesterMixin):
+    """FasterCache tests for Flux Transformer."""
+
+    # Flux is guidance distilled, so we can test at model level without CFG batch handling
+    FASTER_CACHE_CONFIG = {
+        "spatial_attention_block_skip_range": 2,
+        "spatial_attention_timestep_skip_range": (-1, 901),
+        "tensor_format": "BCHW",
+        "is_guidance_distilled": True,
+    }

tests/pipelines/bria_fibo_edit/test_pipeline_bria_fibo_edit.py

@@ -1,192 +0,0 @@
-# Copyright 2024 Bria AI and 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 unittest
-
-import numpy as np
-import torch
-from PIL import Image
-from transformers import AutoTokenizer
-from transformers.models.smollm3.modeling_smollm3 import SmolLM3Config, SmolLM3ForCausalLM
-
-from diffusers import (
-    AutoencoderKLWan,
-    BriaFiboEditPipeline,
-    FlowMatchEulerDiscreteScheduler,
-)
-from diffusers.models.transformers.transformer_bria_fibo import BriaFiboTransformer2DModel
-from tests.pipelines.test_pipelines_common import PipelineTesterMixin
-
-from ...testing_utils import (
-    enable_full_determinism,
-    torch_device,
-)
-
-
-enable_full_determinism()
-
-
-class BriaFiboPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
-    pipeline_class = BriaFiboEditPipeline
-    params = frozenset(["prompt", "height", "width", "guidance_scale"])
-    batch_params = frozenset(["prompt"])
-    test_xformers_attention = False
-    test_layerwise_casting = False
-    test_group_offloading = False
-    supports_dduf = False
-
-    def get_dummy_components(self):
-        torch.manual_seed(0)
-        transformer = BriaFiboTransformer2DModel(
-            patch_size=1,
-            in_channels=16,
-            num_layers=1,
-            num_single_layers=1,
-            attention_head_dim=8,
-            num_attention_heads=2,
-            joint_attention_dim=64,
-            text_encoder_dim=32,
-            pooled_projection_dim=None,
-            axes_dims_rope=[0, 4, 4],
-        )
-
-        vae = AutoencoderKLWan(
-            base_dim=80,
-            decoder_base_dim=128,
-            dim_mult=[1, 2, 4, 4],
-            dropout=0.0,
-            in_channels=12,
-            latents_mean=[0.0] * 16,
-            latents_std=[1.0] * 16,
-            is_residual=True,
-            num_res_blocks=2,
-            out_channels=12,
-            patch_size=2,
-            scale_factor_spatial=16,
-            scale_factor_temporal=4,
-            temperal_downsample=[False, True, True],
-            z_dim=16,
-        )
-        scheduler = FlowMatchEulerDiscreteScheduler()
-        text_encoder = SmolLM3ForCausalLM(SmolLM3Config(hidden_size=32))
-        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
-
-        components = {
-            "scheduler": scheduler,
-            "text_encoder": text_encoder,
-            "tokenizer": tokenizer,
-            "transformer": transformer,
-            "vae": vae,
-        }
-        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="cpu").manual_seed(seed)
-        inputs = {
-            "prompt": '{"text": "A painting of a squirrel eating a burger","edit_instruction": "A painting of a squirrel eating a burger"}',
-            "negative_prompt": "bad, ugly",
-            "generator": generator,
-            "num_inference_steps": 2,
-            "guidance_scale": 5.0,
-            "height": 192,
-            "width": 336,
-            "output_type": "np",
-        }
-        image = Image.new("RGB", (336, 192), (255, 255, 255))
-        inputs["image"] = image
-        return inputs
-
-    @unittest.skip(reason="will not be supported due to dim-fusion")
-    def test_encode_prompt_works_in_isolation(self):
-        pass
-
-    @unittest.skip(reason="Batching is not supported yet")
-    def test_num_images_per_prompt(self):
-        pass
-
-    @unittest.skip(reason="Batching is not supported yet")
-    def test_inference_batch_consistent(self):
-        pass
-
-    @unittest.skip(reason="Batching is not supported yet")
-    def test_inference_batch_single_identical(self):
-        pass
-
-    def test_bria_fibo_different_prompts(self):
-        pipe = self.pipeline_class(**self.get_dummy_components())
-        pipe = pipe.to(torch_device)
-        inputs = self.get_dummy_inputs(torch_device)
-        output_same_prompt = pipe(**inputs).images[0]
-
-        inputs = self.get_dummy_inputs(torch_device)
-        inputs["prompt"] = {"edit_instruction": "a different prompt"}
-        output_different_prompts = pipe(**inputs).images[0]
-
-        max_diff = np.abs(output_same_prompt - output_different_prompts).max()
-        assert max_diff > 1e-6
-
-    def test_image_output_shape(self):
-        pipe = self.pipeline_class(**self.get_dummy_components())
-        pipe = pipe.to(torch_device)
-        inputs = self.get_dummy_inputs(torch_device)
-
-        height_width_pairs = [(32, 32), (64, 64), (32, 64)]
-        for height, width in height_width_pairs:
-            expected_height = height
-            expected_width = width
-
-            inputs.update({"height": height, "width": width})
-            image = pipe(**inputs).images[0]
-            output_height, output_width, _ = image.shape
-            assert (output_height, output_width) == (expected_height, expected_width)
-
-    def test_bria_fibo_edit_mask(self):
-        pipe = self.pipeline_class(**self.get_dummy_components())
-        pipe = pipe.to(torch_device)
-        inputs = self.get_dummy_inputs(torch_device)
-
-        mask = Image.fromarray((np.ones((192, 336)) * 255).astype(np.uint8), mode="L")
-
-        inputs.update({"mask": mask})
-        output = pipe(**inputs).images[0]
-
-        assert output.shape == (192, 336, 3)
-
-    def test_bria_fibo_edit_mask_image_size_mismatch(self):
-        pipe = self.pipeline_class(**self.get_dummy_components())
-        pipe = pipe.to(torch_device)
-        inputs = self.get_dummy_inputs(torch_device)
-
-        mask = Image.fromarray((np.ones((64, 64)) * 255).astype(np.uint8), mode="L")
-
-        inputs.update({"mask": mask})
-        with self.assertRaisesRegex(ValueError, "Mask and image must have the same size"):
-            pipe(**inputs)
-
-    def test_bria_fibo_edit_mask_no_image(self):
-        pipe = self.pipeline_class(**self.get_dummy_components())
-        pipe = pipe.to(torch_device)
-        inputs = self.get_dummy_inputs(torch_device)
-
-        mask = Image.fromarray((np.ones((32, 32)) * 255).astype(np.uint8), mode="L")
-
-        # Remove image from inputs if it's there (it shouldn't be by default from get_dummy_inputs)
-        inputs.pop("image", None)
-        inputs.update({"mask": mask})
-
-        with self.assertRaisesRegex(ValueError, "If mask is provided, image must also be provided"):
-            pipe(**inputs)

tests/testing_utils.py

@@ -38,6 +38,7 @@ from diffusers.utils.import_utils import (
     is_gguf_available,
     is_kernels_available,
     is_note_seq_available,
+    is_nvidia_modelopt_available,
     is_onnx_available,
     is_opencv_available,
     is_optimum_quanto_available,
@@ -130,6 +131,59 @@ def torch_all_close(a, b, *args, **kwargs):
     return True
 
 
+def assert_tensors_close(
+    actual: "torch.Tensor",
+    expected: "torch.Tensor",
+    atol: float = 1e-5,
+    rtol: float = 1e-5,
+    msg: str = "",
+) -> None:
+    """
+    Assert that two tensors are close within tolerance.
+
+    Uses the same formula as torch.allclose: |actual - expected| <= atol + rtol * |expected|
+    Provides concise, actionable error messages without dumping full tensors.
+
+    Args:
+        actual: The actual tensor from the computation.
+        expected: The expected tensor to compare against.
+        atol: Absolute tolerance.
+        rtol: Relative tolerance.
+        msg: Optional message prefix for the assertion error.
+
+    Raises:
+        AssertionError: If tensors have different shapes or values exceed tolerance.
+
+    Example:
+        >>> assert_tensors_close(output, expected_output, atol=1e-5, rtol=1e-5, msg="Forward pass")
+    """
+    if not is_torch_available():
+        raise ValueError("PyTorch needs to be installed to use this function.")
+
+    if actual.shape != expected.shape:
+        raise AssertionError(f"{msg} Shape mismatch: actual {actual.shape} vs expected {expected.shape}")
+
+    if not torch.allclose(actual, expected, atol=atol, rtol=rtol):
+        abs_diff = (actual - expected).abs()
+        max_diff = abs_diff.max().item()
+
+        flat_idx = abs_diff.argmax().item()
+        max_idx = tuple(torch.unravel_index(torch.tensor(flat_idx), actual.shape).tolist())
+
+        threshold = atol + rtol * expected.abs()
+        mismatched = (abs_diff > threshold).sum().item()
+        total = actual.numel()
+
+        raise AssertionError(
+            f"{msg}\n"
+            f"Tensors not close! Mismatched elements: {mismatched}/{total} ({100 * mismatched / total:.1f}%)\n"
+            f"  Max diff: {max_diff:.6e} at index {max_idx}\n"
+            f"  Actual:   {actual.flatten()[flat_idx].item():.6e}\n"
+            f"  Expected: {expected.flatten()[flat_idx].item():.6e}\n"
+            f"  atol: {atol:.6e}, rtol: {rtol:.6e}"
+        )
+
+
 def numpy_cosine_similarity_distance(a, b):
     similarity = np.dot(a, b) / (norm(a) * norm(b))
     distance = 1.0 - similarity.mean()
@@ -241,7 +295,6 @@ def parse_flag_from_env(key, default=False):
 
 _run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False)
 _run_nightly_tests = parse_flag_from_env("RUN_NIGHTLY", default=False)
-_run_compile_tests = parse_flag_from_env("RUN_COMPILE", default=False)
 
 
 def floats_tensor(shape, scale=1.0, rng=None, name=None):
@@ -282,12 +335,155 @@ def nightly(test_case):
 
 def is_torch_compile(test_case):
     """
-    Decorator marking a test that runs compile tests in the diffusers CI.
-
-    Compile tests are skipped by default. Set the RUN_COMPILE environment variable to a truthy value to run them.
-
+    Decorator marking a test as a torch.compile test. These tests can be filtered using:
+        pytest -m "not compile" to skip
+        pytest -m compile to run only these tests
     """
-    return pytest.mark.skipif(not _run_compile_tests, reason="test is torch compile")(test_case)
+    return pytest.mark.compile(test_case)
+
+
+def is_single_file(test_case):
+    """
+    Decorator marking a test as a single file loading test. These tests can be filtered using:
+        pytest -m "not single_file" to skip
+        pytest -m single_file to run only these tests
+    """
+    return pytest.mark.single_file(test_case)
+
+
+def is_lora(test_case):
+    """
+    Decorator marking a test as a LoRA test. These tests can be filtered using:
+        pytest -m "not lora" to skip
+        pytest -m lora to run only these tests
+    """
+    return pytest.mark.lora(test_case)
+
+
+def is_ip_adapter(test_case):
+    """
+    Decorator marking a test as an IP Adapter test. These tests can be filtered using:
+        pytest -m "not ip_adapter" to skip
+        pytest -m ip_adapter to run only these tests
+    """
+    return pytest.mark.ip_adapter(test_case)
+
+
+def is_training(test_case):
+    """
+    Decorator marking a test as a training test. These tests can be filtered using:
+        pytest -m "not training" to skip
+        pytest -m training to run only these tests
+    """
+    return pytest.mark.training(test_case)
+
+
+def is_attention(test_case):
+    """
+    Decorator marking a test as an attention test. These tests can be filtered using:
+        pytest -m "not attention" to skip
+        pytest -m attention to run only these tests
+    """
+    return pytest.mark.attention(test_case)
+
+
+def is_memory(test_case):
+    """
+    Decorator marking a test as a memory optimization test. These tests can be filtered using:
+        pytest -m "not memory" to skip
+        pytest -m memory to run only these tests
+    """
+    return pytest.mark.memory(test_case)
+
+
+def is_cpu_offload(test_case):
+    """
+    Decorator marking a test as a CPU offload test. These tests can be filtered using:
+        pytest -m "not cpu_offload" to skip
+        pytest -m cpu_offload to run only these tests
+    """
+    return pytest.mark.cpu_offload(test_case)
+
+
+def is_group_offload(test_case):
+    """
+    Decorator marking a test as a group offload test. These tests can be filtered using:
+        pytest -m "not group_offload" to skip
+        pytest -m group_offload to run only these tests
+    """
+    return pytest.mark.group_offload(test_case)
+
+
+def is_quantization(test_case):
+    """
+    Decorator marking a test as a quantization test. These tests can be filtered using:
+        pytest -m "not quantization" to skip
+        pytest -m quantization to run only these tests
+    """
+    return pytest.mark.quantization(test_case)
+
+
+def is_bitsandbytes(test_case):
+    """
+    Decorator marking a test as a BitsAndBytes quantization test. These tests can be filtered using:
+        pytest -m "not bitsandbytes" to skip
+        pytest -m bitsandbytes to run only these tests
+    """
+    return pytest.mark.bitsandbytes(test_case)
+
+
+def is_quanto(test_case):
+    """
+    Decorator marking a test as a Quanto quantization test. These tests can be filtered using:
+        pytest -m "not quanto" to skip
+        pytest -m quanto to run only these tests
+    """
+    return pytest.mark.quanto(test_case)
+
+
+def is_torchao(test_case):
+    """
+    Decorator marking a test as a TorchAO quantization test. These tests can be filtered using:
+        pytest -m "not torchao" to skip
+        pytest -m torchao to run only these tests
+    """
+    return pytest.mark.torchao(test_case)
+
+
+def is_gguf(test_case):
+    """
+    Decorator marking a test as a GGUF quantization test. These tests can be filtered using:
+        pytest -m "not gguf" to skip
+        pytest -m gguf to run only these tests
+    """
+    return pytest.mark.gguf(test_case)
+
+
+def is_modelopt(test_case):
+    """
+    Decorator marking a test as a NVIDIA ModelOpt quantization test. These tests can be filtered using:
+        pytest -m "not modelopt" to skip
+        pytest -m modelopt to run only these tests
+    """
+    return pytest.mark.modelopt(test_case)
+
+
+def is_context_parallel(test_case):
+    """
+    Decorator marking a test as a context parallel inference test. These tests can be filtered using:
+        pytest -m "not context_parallel" to skip
+        pytest -m context_parallel to run only these tests
+    """
+    return pytest.mark.context_parallel(test_case)
+
+
+def is_cache(test_case):
+    """
+    Decorator marking a test as a cache test. These tests can be filtered using:
+        pytest -m "not cache" to skip
+        pytest -m cache to run only these tests
+    """
+    return pytest.mark.cache(test_case)
 
 
 def require_torch(test_case):
@@ -650,6 +846,19 @@ def require_kernels_version_greater_or_equal(kernels_version):
     return decorator
 
 
+def require_modelopt_version_greater_or_equal(modelopt_version):
+    def decorator(test_case):
+        correct_nvidia_modelopt_version = is_nvidia_modelopt_available() and version.parse(
+            version.parse(importlib.metadata.version("modelopt")).base_version
+        ) >= version.parse(modelopt_version)
+        return pytest.mark.skipif(
+            not correct_nvidia_modelopt_version,
+            reason=f"Test requires modelopt with version greater than {modelopt_version}.",
+        )(test_case)
+
+    return decorator
+
+
 def deprecate_after_peft_backend(test_case):
     """
     Decorator marking a test that will be skipped after PEFT backend

utils/generate_model_tests.py

@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2025 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.
+
+"""
+Utility script to generate test suites for diffusers model classes.
+
+Usage:
+    python utils/generate_model_tests.py src/diffusers/models/transformers/transformer_flux.py
+
+This will analyze the model file and generate a test file with appropriate
+test classes based on the model's mixins and attributes.
+"""
+
+import argparse
+import ast
+import sys
+from pathlib import Path
+
+
+MIXIN_TO_TESTER = {
+    "ModelMixin": "ModelTesterMixin",
+    "PeftAdapterMixin": "LoraTesterMixin",
+}
+
+ATTRIBUTE_TO_TESTER = {
+    "_cp_plan": "ContextParallelTesterMixin",
+    "_supports_gradient_checkpointing": "TrainingTesterMixin",
+}
+
+ALWAYS_INCLUDE_TESTERS = [
+    "ModelTesterMixin",
+    "MemoryTesterMixin",
+    "TorchCompileTesterMixin",
+]
+
+# Attention-related class names that indicate the model uses attention
+ATTENTION_INDICATORS = {
+    "AttentionMixin",
+    "AttentionModuleMixin",
+}
+
+OPTIONAL_TESTERS = [
+    # Quantization testers
+    ("BitsAndBytesTesterMixin", "bnb"),
+    ("QuantoTesterMixin", "quanto"),
+    ("TorchAoTesterMixin", "torchao"),
+    ("GGUFTesterMixin", "gguf"),
+    ("ModelOptTesterMixin", "modelopt"),
+    # Quantization compile testers
+    ("BitsAndBytesCompileTesterMixin", "bnb_compile"),
+    ("QuantoCompileTesterMixin", "quanto_compile"),
+    ("TorchAoCompileTesterMixin", "torchao_compile"),
+    ("GGUFCompileTesterMixin", "gguf_compile"),
+    ("ModelOptCompileTesterMixin", "modelopt_compile"),
+    # Cache testers
+    ("PyramidAttentionBroadcastTesterMixin", "pab_cache"),
+    ("FirstBlockCacheTesterMixin", "fbc_cache"),
+    ("FasterCacheTesterMixin", "faster_cache"),
+    # Other testers
+    ("SingleFileTesterMixin", "single_file"),
+    ("IPAdapterTesterMixin", "ip_adapter"),
+]
+
+
+class ModelAnalyzer(ast.NodeVisitor):
+    def __init__(self):
+        self.model_classes = []
+        self.current_class = None
+        self.imports = set()
+
+    def visit_Import(self, node: ast.Import):
+        for alias in node.names:
+            self.imports.add(alias.name.split(".")[-1])
+        self.generic_visit(node)
+
+    def visit_ImportFrom(self, node: ast.ImportFrom):
+        for alias in node.names:
+            self.imports.add(alias.name)
+        self.generic_visit(node)
+
+    def visit_ClassDef(self, node: ast.ClassDef):
+        base_names = []
+        for base in node.bases:
+            if isinstance(base, ast.Name):
+                base_names.append(base.id)
+            elif isinstance(base, ast.Attribute):
+                base_names.append(base.attr)
+
+        if "ModelMixin" in base_names:
+            class_info = {
+                "name": node.name,
+                "bases": base_names,
+                "attributes": {},
+                "has_forward": False,
+                "init_params": [],
+            }
+
+            for item in node.body:
+                if isinstance(item, ast.Assign):
+                    for target in item.targets:
+                        if isinstance(target, ast.Name):
+                            attr_name = target.id
+                            if attr_name.startswith("_"):
+                                class_info["attributes"][attr_name] = self._get_value(item.value)
+
+                elif isinstance(item, ast.FunctionDef):
+                    if item.name == "forward":
+                        class_info["has_forward"] = True
+                        class_info["forward_params"] = self._extract_func_params(item)
+                    elif item.name == "__init__":
+                        class_info["init_params"] = self._extract_func_params(item)
+
+            self.model_classes.append(class_info)
+
+        self.generic_visit(node)
+
+    def _extract_func_params(self, func_node: ast.FunctionDef) -> list[dict]:
+        params = []
+        args = func_node.args
+
+        num_defaults = len(args.defaults)
+        num_args = len(args.args)
+        first_default_idx = num_args - num_defaults
+
+        for i, arg in enumerate(args.args):
+            if arg.arg == "self":
+                continue
+
+            param_info = {"name": arg.arg, "type": None, "default": None}
+
+            if arg.annotation:
+                param_info["type"] = self._get_annotation_str(arg.annotation)
+
+            default_idx = i - first_default_idx
+            if default_idx >= 0 and default_idx < len(args.defaults):
+                param_info["default"] = self._get_value(args.defaults[default_idx])
+
+            params.append(param_info)
+
+        return params
+
+    def _get_annotation_str(self, node) -> str:
+        if isinstance(node, ast.Name):
+            return node.id
+        elif isinstance(node, ast.Constant):
+            return repr(node.value)
+        elif isinstance(node, ast.Subscript):
+            base = self._get_annotation_str(node.value)
+            if isinstance(node.slice, ast.Tuple):
+                args = ", ".join(self._get_annotation_str(el) for el in node.slice.elts)
+            else:
+                args = self._get_annotation_str(node.slice)
+            return f"{base}[{args}]"
+        elif isinstance(node, ast.Attribute):
+            return f"{self._get_annotation_str(node.value)}.{node.attr}"
+        elif isinstance(node, ast.BinOp) and isinstance(node.op, ast.BitOr):
+            left = self._get_annotation_str(node.left)
+            right = self._get_annotation_str(node.right)
+            return f"{left} | {right}"
+        elif isinstance(node, ast.Tuple):
+            return ", ".join(self._get_annotation_str(el) for el in node.elts)
+        return "Any"
+
+    def _get_value(self, node):
+        if isinstance(node, ast.Constant):
+            return node.value
+        elif isinstance(node, ast.Name):
+            if node.id == "None":
+                return None
+            elif node.id == "True":
+                return True
+            elif node.id == "False":
+                return False
+            return node.id
+        elif isinstance(node, ast.List):
+            return [self._get_value(el) for el in node.elts]
+        elif isinstance(node, ast.Dict):
+            return {self._get_value(k): self._get_value(v) for k, v in zip(node.keys, node.values)}
+        return "<complex>"
+
+
+def analyze_model_file(filepath: str) -> tuple[list[dict], set[str]]:
+    with open(filepath) as f:
+        source = f.read()
+
+    tree = ast.parse(source)
+    analyzer = ModelAnalyzer()
+    analyzer.visit(tree)
+
+    return analyzer.model_classes, analyzer.imports
+
+
+def determine_testers(model_info: dict, include_optional: list[str], imports: set[str]) -> list[str]:
+    testers = list(ALWAYS_INCLUDE_TESTERS)
+
+    for base in model_info["bases"]:
+        if base in MIXIN_TO_TESTER:
+            tester = MIXIN_TO_TESTER[base]
+            if tester not in testers:
+                testers.append(tester)
+
+    for attr, tester in ATTRIBUTE_TO_TESTER.items():
+        if attr in model_info["attributes"]:
+            value = model_info["attributes"][attr]
+            if value is not None and value is not False:
+                if tester not in testers:
+                    testers.append(tester)
+
+    if "_cp_plan" in model_info["attributes"] and model_info["attributes"]["_cp_plan"] is not None:
+        if "ContextParallelTesterMixin" not in testers:
+            testers.append("ContextParallelTesterMixin")
+
+    # Include AttentionTesterMixin if the model imports attention-related classes
+    if imports & ATTENTION_INDICATORS:
+        testers.append("AttentionTesterMixin")
+
+    for tester, flag in OPTIONAL_TESTERS:
+        if flag in include_optional:
+            if tester not in testers:
+                testers.append(tester)
+
+    return testers
+
+
+def generate_config_class(model_info: dict, model_name: str) -> str:
+    class_name = f"{model_name}TesterConfig"
+    model_class = model_info["name"]
+    forward_params = model_info.get("forward_params", [])
+    init_params = model_info.get("init_params", [])
+
+    lines = [
+        f"class {class_name}:",
+        "    @property",
+        "    def model_class(self):",
+        f"        return {model_class}",
+        "",
+        "    @property",
+        "    def pretrained_model_name_or_path(self):",
+        '        return ""  # TODO: Set Hub repository ID',
+        "",
+        "    @property",
+        "    def pretrained_model_kwargs(self):",
+        '        return {"subfolder": "transformer"}',
+        "",
+        "    @property",
+        "    def generator(self):",
+        '        return torch.Generator("cpu").manual_seed(0)',
+        "",
+        "    def get_init_dict(self) -> dict[str, int | list[int]]:",
+    ]
+
+    if init_params:
+        lines.append("        # __init__ parameters:")
+        for param in init_params:
+            type_str = f": {param['type']}" if param["type"] else ""
+            default_str = f" = {param['default']}" if param["default"] is not None else ""
+            lines.append(f"        #   {param['name']}{type_str}{default_str}")
+
+    lines.extend(
+        [
+            "        return {}",
+            "",
+            "    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:",
+        ]
+    )
+
+    if forward_params:
+        lines.append("        # forward() parameters:")
+        for param in forward_params:
+            type_str = f": {param['type']}" if param["type"] else ""
+            default_str = f" = {param['default']}" if param["default"] is not None else ""
+            lines.append(f"        #   {param['name']}{type_str}{default_str}")
+
+    lines.extend(
+        [
+            "        # TODO: Fill in dummy inputs",
+            "        return {}",
+            "",
+            "    @property",
+            "    def input_shape(self) -> tuple[int, ...]:",
+            "        return (1, 1)",
+            "",
+            "    @property",
+            "    def output_shape(self) -> tuple[int, ...]:",
+            "        return (1, 1)",
+        ]
+    )
+
+    return "\n".join(lines)
+
+
+def generate_test_class(model_name: str, config_class: str, tester: str) -> str:
+    tester_short = tester.replace("TesterMixin", "")
+    class_name = f"Test{model_name}{tester_short}"
+
+    lines = [f"class {class_name}({config_class}, {tester}):"]
+
+    if tester == "TorchCompileTesterMixin":
+        lines.extend(
+            [
+                "    @property",
+                "    def different_shapes_for_compilation(self):",
+                "        return [(4, 4), (4, 8), (8, 8)]",
+                "",
+                "    def get_dummy_inputs(self, height: int = 4, width: int = 4) -> dict[str, torch.Tensor]:",
+                "        # TODO: Implement dynamic input generation",
+                "        return {}",
+            ]
+        )
+    elif tester == "IPAdapterTesterMixin":
+        lines.extend(
+            [
+                "    @property",
+                "    def ip_adapter_processor_cls(self):",
+                "        return None  # TODO: Set processor class",
+                "",
+                "    def modify_inputs_for_ip_adapter(self, model, inputs_dict):",
+                "        # TODO: Add IP adapter image embeds to inputs",
+                "        return inputs_dict",
+                "",
+                "    def create_ip_adapter_state_dict(self, model):",
+                "        # TODO: Create IP adapter state dict",
+                "        return {}",
+            ]
+        )
+    elif tester == "SingleFileTesterMixin":
+        lines.extend(
+            [
+                "    @property",
+                "    def ckpt_path(self):",
+                '        return ""  # TODO: Set checkpoint path',
+                "",
+                "    @property",
+                "    def alternate_ckpt_paths(self):",
+                "        return []",
+                "",
+                "    @property",
+                "    def pretrained_model_name_or_path(self):",
+                '        return ""  # TODO: Set Hub repository ID',
+            ]
+        )
+    elif tester == "GGUFTesterMixin":
+        lines.extend(
+            [
+                "    @property",
+                "    def gguf_filename(self):",
+                '        return ""  # TODO: Set GGUF filename',
+                "",
+                "    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:",
+                "        # TODO: Override with larger inputs for quantization tests",
+                "        return {}",
+            ]
+        )
+    elif tester in ["BitsAndBytesTesterMixin", "QuantoTesterMixin", "TorchAoTesterMixin", "ModelOptTesterMixin"]:
+        lines.extend(
+            [
+                "    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:",
+                "        # TODO: Override with larger inputs for quantization tests",
+                "        return {}",
+            ]
+        )
+    elif tester in [
+        "BitsAndBytesCompileTesterMixin",
+        "QuantoCompileTesterMixin",
+        "TorchAoCompileTesterMixin",
+        "ModelOptCompileTesterMixin",
+    ]:
+        lines.extend(
+            [
+                "    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:",
+                "        # TODO: Override with larger inputs for quantization compile tests",
+                "        return {}",
+            ]
+        )
+    elif tester == "GGUFCompileTesterMixin":
+        lines.extend(
+            [
+                "    @property",
+                "    def gguf_filename(self):",
+                '        return ""  # TODO: Set GGUF filename',
+                "",
+                "    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:",
+                "        # TODO: Override with larger inputs for quantization compile tests",
+                "        return {}",
+            ]
+        )
+    elif tester in [
+        "PyramidAttentionBroadcastTesterMixin",
+        "FirstBlockCacheTesterMixin",
+        "FasterCacheTesterMixin",
+    ]:
+        lines.append("    pass")
+    elif tester == "LoraHotSwappingForModelTesterMixin":
+        lines.extend(
+            [
+                "    @property",
+                "    def different_shapes_for_compilation(self):",
+                "        return [(4, 4), (4, 8), (8, 8)]",
+                "",
+                "    def get_dummy_inputs(self, height: int = 4, width: int = 4) -> dict[str, torch.Tensor]:",
+                "        # TODO: Implement dynamic input generation",
+                "        return {}",
+            ]
+        )
+    else:
+        lines.append("    pass")
+
+    return "\n".join(lines)
+
+
+def generate_test_file(model_info: dict, model_filepath: str, include_optional: list[str], imports: set[str]) -> str:
+    model_name = model_info["name"].replace("2DModel", "").replace("3DModel", "").replace("Model", "")
+    testers = determine_testers(model_info, include_optional, imports)
+    tester_imports = sorted(set(testers) - {"LoraHotSwappingForModelTesterMixin"})
+
+    lines = [
+        "# coding=utf-8",
+        "# Copyright 2025 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 torch",
+        "",
+        f"from diffusers import {model_info['name']}",
+        "from diffusers.utils.torch_utils import randn_tensor",
+        "",
+        "from ...testing_utils import enable_full_determinism, torch_device",
+    ]
+
+    if "LoraTesterMixin" in testers:
+        lines.append("from ..test_modeling_common import LoraHotSwappingForModelTesterMixin")
+
+    lines.extend(
+        [
+            "from ..testing_utils import (",
+            *[f"    {tester}," for tester in sorted(tester_imports)],
+            ")",
+            "",
+            "",
+            "enable_full_determinism()",
+            "",
+            "",
+        ]
+    )
+
+    config_class = f"{model_name}TesterConfig"
+    lines.append(generate_config_class(model_info, model_name))
+    lines.append("")
+    lines.append("")
+
+    for tester in testers:
+        lines.append(generate_test_class(model_name, config_class, tester))
+        lines.append("")
+        lines.append("")
+
+    if "LoraTesterMixin" in testers:
+        lines.append(generate_test_class(model_name, config_class, "LoraHotSwappingForModelTesterMixin"))
+        lines.append("")
+        lines.append("")
+
+    return "\n".join(lines).rstrip() + "\n"
+
+
+def get_test_output_path(model_filepath: str) -> str:
+    path = Path(model_filepath)
+    model_filename = path.stem
+
+    if "transformers" in path.parts:
+        return f"tests/models/transformers/test_models_{model_filename}.py"
+    elif "unets" in path.parts:
+        return f"tests/models/unets/test_models_{model_filename}.py"
+    elif "autoencoders" in path.parts:
+        return f"tests/models/autoencoders/test_models_{model_filename}.py"
+    else:
+        return f"tests/models/test_models_{model_filename}.py"
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Generate test suite for a diffusers model class")
+    parser.add_argument(
+        "model_filepath",
+        type=str,
+        help="Path to the model file (e.g., src/diffusers/models/transformers/transformer_flux.py)",
+    )
+    parser.add_argument(
+        "--output", "-o", type=str, default=None, help="Output file path (default: auto-generated based on model path)"
+    )
+    parser.add_argument(
+        "--include",
+        "-i",
+        type=str,
+        nargs="*",
+        default=[],
+        choices=[
+            "bnb",
+            "quanto",
+            "torchao",
+            "gguf",
+            "modelopt",
+            "bnb_compile",
+            "quanto_compile",
+            "torchao_compile",
+            "gguf_compile",
+            "modelopt_compile",
+            "pab_cache",
+            "fbc_cache",
+            "faster_cache",
+            "single_file",
+            "ip_adapter",
+            "all",
+        ],
+        help="Optional testers to include",
+    )
+    parser.add_argument(
+        "--class-name",
+        "-c",
+        type=str,
+        default=None,
+        help="Specific model class to generate tests for (default: first model class found)",
+    )
+    parser.add_argument("--dry-run", action="store_true", help="Print generated code without writing to file")
+
+    args = parser.parse_args()
+
+    if not Path(args.model_filepath).exists():
+        print(f"Error: File not found: {args.model_filepath}", file=sys.stderr)
+        sys.exit(1)
+
+    model_classes, imports = analyze_model_file(args.model_filepath)
+
+    if not model_classes:
+        print(f"Error: No model classes found in {args.model_filepath}", file=sys.stderr)
+        sys.exit(1)
+
+    if args.class_name:
+        model_info = next((m for m in model_classes if m["name"] == args.class_name), None)
+        if not model_info:
+            available = [m["name"] for m in model_classes]
+            print(f"Error: Class '{args.class_name}' not found. Available: {available}", file=sys.stderr)
+            sys.exit(1)
+    else:
+        model_info = model_classes[0]
+        if len(model_classes) > 1:
+            print(f"Multiple model classes found, using: {model_info['name']}", file=sys.stderr)
+            print("Use --class-name to specify a different class", file=sys.stderr)
+
+    include_optional = args.include
+    if "all" in include_optional:
+        include_optional = [flag for _, flag in OPTIONAL_TESTERS]
+
+    generated_code = generate_test_file(model_info, args.model_filepath, include_optional, imports)
+
+    if args.dry_run:
+        print(generated_code)
+    else:
+        output_path = args.output or get_test_output_path(args.model_filepath)
+        output_dir = Path(output_path).parent
+        output_dir.mkdir(parents=True, exist_ok=True)
+
+        with open(output_path, "w") as f:
+            f.write(generated_code)
+
+        print(f"Generated test file: {output_path}")
+        print(f"Model class: {model_info['name']}")
+        print(f"Detected attributes: {list(model_info['attributes'].keys())}")
+
+
+if __name__ == "__main__":
+    main()