Welcome to the ExecuTorch documentation! This README.md will provide an overview of the ExecuTorch docs and its features, as well as instructions on how to contribute and build locally.
We are using sphinx with myst_parser, sphinx-gallery, and sphinx_design in this documentation set.
We support both .rst and .md files but prefer the content to be authored in .md as much as possible.
Documentation dependancies are stored in .ci/docker/requirements-ci.txt.
To build the documentation locally:
Clone the executorch repo to your machine.
Switch to the docs/ directory.
Start a virtual environment:
virtualenv venv source venv/bin/activate
Install dependancies:
pip3 install -r ../.ci/docker/requirements-ci.txt
Build the documentation set:
make html
This should build both documentation and tutorials. The build will be placed in the _build directory.
You can preview locally by using sphinx-serve. To install sphinx-serve, run: pip3 install sphinx-serve. To serve your documentation:
sphinx-serve -b _build
Open http://0.0.0.0:8081/ in your browser to preview your updated documentation.
You can use custom variables in your .md and .rst files. The variables take their values from the files listed in the ./.ci/docker/ci_commit_pins/ directory. For example, to insert a variable that specifies the latest Buck2 version, use the following syntax:
The current version of Buck2 is ${executorch_version:buck2}.
This will result in the following output:
We have the following custom variables defined in this docset:
${executorch_version:buck2}${executorch_version:nightly}${executorch_version:pytorch}${executorch_version:vision}${executorch_version:audio}You can use the variables in both regular text and code blocks.
You might want to include some of the README.md files from various directories in this repositories in your documentation build. To do that, create an .md file and use the {include} directive to insert your .md files. Example:
```{include} ../README.md
NOTE: Many README.md files are written as placeholders with limited information provided. Some of that content you might want to keep in the repository rather than on the website. If you still want to add it, make sure to check the content for accuracy, structure, and overall quality.
Use the PyTorch contributing guidelines to contribute to the documentation.
In addition to that, see Markdown in Sphinx Tips and Tricks for tips on how to author high-quality markdown pages with Myst Parser.
On a high level (will go into detail later), we use Sphinx to generate both Python and C++ documentation in the form of HTML pages.
We generate Python API documentation through Sphinx, bootstrapping Pytorch's Sphinx theme for a cohesive look with the existing Pytorch API documentation.
The setup for Python documentation lies within source_py/. To set up Sphinx, a conf.py configuration file is required. We can specify ways to generate documentation here. Specifically, the most important/relevant parts are:
Make sure to add a path to the directory above the directory you're trying to generate documentation for. For example, since we want to generate documenation for the executorch/ directory. This tells Sphinx where to find the code to generate docs for.
extensions contains extension modules. For auto-generating APIs, make sure to include sphinx.ext.autodoc.
autodoc_mock_imports is where you put imports that Sphinx is unable to access. Sphinx runs your code in order to autogenerate the docs, so for any libraries that it unable to access due to it being outside of the directory, or containing c++ bindings, we need to specify it in the autodoc_mock_imports list. You can see what modules Sphinx is confused by when you run into importing errors when generating docs.
Additionally, RST files are needed in order to specify the structure of the auto-generated pages and to tell Sphinx what modules to generate documentation for. To auto-generate APIs for a specific module, the automodule tag is needed to tell Sphinx what specific module to document. For example, if we wanted a page to display auto-generated documentation for everything in executorch/exir/__init__.py, the RST file would look something like the following:
executorch.exir ======================= .. automodule:: executorch.exir :members: :undoc-members: :show-inheritance:
These separate RST files should all be linked together, with the initial landing page under index.rst. A sample of this structure can be found in source_py/.
A diagram of how the files work together: 
To view your changes on the executorch website, you can follow the same steps listed in the “General Documentation” section.
To view just the auto-generated pages:
cd executorch/docs/sphinx-build -M html source_py sphinxbuild_py to build Sphinx and generate APIs any packages.python3 -m http.server 8000 --directory sphinxbuild_py/html to view your HTML files at localhost:8000.Following Pytorch's way of generating C++ documentation, we generate C++ API documentation through Doxygen, which is then converted into Sphinx using Breathe and Exhale.
Specifically, we use Doxygen to generate C++ documentation in the form of XML files, and through configs set in Sphinx's conf.py file, we use Breathe and Exhale to use the XML files and generate RST files which are then used to generate HTML files.
To configure Doxygen, we can run doxygen -g in the root of our repository (ex. source_cpp) which will generate a Doxyfile containing configurations for generating c++ documentation. Specifically, the most important/relevant parts are:
OUTPUT_DIRECTORY specifies where to output the auto-generated XML filesINPUT specifies which files to generate documenation forGENERATE_XML = YESFollowing PyTorch's conf.py file, we can set up our own conf.py file to take in the directory in which we generated the XML files, and output HTML to another directory. A sample of this structure can be found in source_cpp/.
A diagram of how the files work together: 
To view the C++ API documentation locally, run:
cd executorch/docs/sphinx-build -M html source_cpp sphinxbuild_cpppython3 -m http.server 8000 --directory sphinxbuild_py/html to view your HTML files at localhost:8000.