The PyO3 user guide

Welcome to the PyO3 user guide! This book is a companion to PyO3's API docs. It contains examples and documentation to explain all of PyO3's use cases in detail.

The rough order of material in this user guide is as follows:

1. Getting started
2. Wrapping Rust code for use from Python
3. How to use Python code from Rust
4. Remaining topics which go into advanced concepts in detail

Please choose from the chapters on the left to jump to individual topics, or continue below to start with PyO3's README.

PyO3

Rust bindings for Python, including tools for creating native Python extension modules. Running and interacting with Python code from a Rust binary is also supported.

• User Guide: stable | main

• API Documentation: stable | main

Usage

PyO3 supports the following software versions:

• Python 3.7 and up (CPython, PyPy, and GraalPy)
• Rust 1.56 and up

You can use PyO3 to write a native Python module in Rust, or to embed Python in a Rust binary. The following sections explain each of these in turn.

Using Rust from Python

PyO3 can be used to generate a native Python module. The easiest way to try this out for the first time is to use maturin. maturin is a tool for building and publishing Rust-based Python packages with minimal configuration. The following steps install maturin, use it to generate and build a new Python package, and then launch Python to import and execute a function from the package.

First, follow the commands below to create a new directory containing a new Python virtualenv, and install maturin into the virtualenv using Python's package manager, pip:

# (replace string_sum with the desired package name)
$ mkdir string_sum
$ cd string_sum
$ python -m venv .env
$ source .env/bin/activate
$ pip install maturin

Still inside this string_sum directory, now run maturin init. This will generate the new package source. When given the choice of bindings to use, select pyo3 bindings:

$ maturin init
✔ 🤷 What kind of bindings to use? · pyo3
  ✨ Done! New project created string_sum

The most important files generated by this command are Cargo.toml and lib.rs, which will look roughly like the following:

Cargo.toml

[package]
name = "string_sum"
version = "0.1.0"
edition = "2021"

[lib]
# The name of the native library. This is the name which will be used in Python to import the
# library (i.e. `import string_sum`). If you change this, you must also change the name of the
# `#[pymodule]` in `src/lib.rs`.
name = "string_sum"
# "cdylib" is necessary to produce a shared library for Python to import from.
#
# Downstream Rust code (including code in `bin/`, `examples/`, and `tests/`) will not be able
# to `use string_sum;` unless the "rlib" or "lib" crate type is also included, e.g.:
# crate-type = ["cdylib", "rlib"]
crate-type = ["cdylib"]

[dependencies]
pyo3 = { version = "0.21.2", features = ["extension-module"] }

src/lib.rs

use pyo3::prelude::*;

/// Formats the sum of two numbers as string.
#[pyfunction]
fn sum_as_string(a: usize, b: usize) -> PyResult<String> {
    Ok((a + b).to_string())
}

/// A Python module implemented in Rust. The name of this function must match
/// the `lib.name` setting in the `Cargo.toml`, else Python will not be able to
/// import the module.
#[pymodule]
fn string_sum(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(sum_as_string, m)?)?;
    Ok(())
}

Finally, run maturin develop. This will build the package and install it into the Python virtualenv previously created and activated. The package is then ready to be used from python:

$ maturin develop
# lots of progress output as maturin runs the compilation...
$ python
>>> import string_sum
>>> string_sum.sum_as_string(5, 20)
'25'

To make changes to the package, just edit the Rust source code and then re-run maturin develop to recompile.

To run this all as a single copy-and-paste, use the bash script below (replace string_sum in the first command with the desired package name):

mkdir string_sum && cd "$_"
python -m venv .env
source .env/bin/activate
pip install maturin
maturin init --bindings pyo3
maturin develop

If you want to be able to run cargo test or use this project in a Cargo workspace and are running into linker issues, there are some workarounds in the FAQ.

As well as with maturin, it is possible to build using setuptools-rust or manually. Both offer more flexibility than maturin but require more configuration to get started.

Using Python from Rust

To embed Python into a Rust binary, you need to ensure that your Python installation contains a shared library. The following steps demonstrate how to ensure this (for Ubuntu), and then give some example code which runs an embedded Python interpreter.

To install the Python shared library on Ubuntu:

sudo apt install python3-dev

To install the Python shared library on RPM based distributions (e.g. Fedora, Red Hat, SuSE), install the python3-devel package.

Start a new project with cargo new and add pyo3 to the Cargo.toml like this:

[dependencies.pyo3]
version = "0.21.2"
features = ["auto-initialize"]

Example program displaying the value of sys.version and the current user name:

use pyo3::prelude::*;
use pyo3::types::IntoPyDict;

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let sys = py.import_bound("sys")?;
        let version: String = sys.getattr("version")?.extract()?;

        let locals = [("os", py.import_bound("os")?)].into_py_dict_bound(py);
        let code = "os.getenv('USER') or os.getenv('USERNAME') or 'Unknown'";
        let user: String = py.eval_bound(code, None, Some(&locals))?.extract()?;

        println!("Hello {}, I'm Python {}", user, version);
        Ok(())
    })
}

The guide has a section with lots of examples about this topic.

Tools and libraries

• maturin Build and publish crates with pyo3, rust-cpython or cffi bindings as well as rust binaries as python packages
• setuptools-rust Setuptools plugin for Rust support.
• pyo3-built Simple macro to expose metadata obtained with the built crate as a PyDict
• rust-numpy Rust binding of NumPy C-API
• dict-derive Derive FromPyObject to automatically transform Python dicts into Rust structs
• pyo3-log Bridge from Rust to Python logging
• pythonize Serde serializer for converting Rust objects to JSON-compatible Python objects
• pyo3-asyncio Utilities for working with Python's Asyncio library and async functions
• rustimport Directly import Rust files or crates from Python, without manual compilation step. Provides pyo3 integration by default and generates pyo3 binding code automatically.

Examples

• autopy A simple, cross-platform GUI automation library for Python and Rust.
  • Contains an example of building wheels on TravisCI and appveyor using cibuildwheel
• ballista-python A Python library that binds to Apache Arrow distributed query engine Ballista.
• bed-reader Read and write the PLINK BED format, simply and efficiently.
  • Shows Rayon/ndarray::parallel (including capturing errors, controlling thread num), Python types to Rust generics, Github Actions
• cryptography Python cryptography library with some functionality in Rust.
• css-inline CSS inlining for Python implemented in Rust.
• datafusion-python A Python library that binds to Apache Arrow in-memory query engine DataFusion.
• deltalake-python Native Delta Lake Python binding based on delta-rs with Pandas integration.
• fastbloom A fast bloom filter | counting bloom filter implemented by Rust for Rust and Python!
• fastuuid Python bindings to Rust's UUID library.
• feos Lightning fast thermodynamic modeling in Rust with fully developed Python interface.
• forust A lightweight gradient boosted decision tree library written in Rust.
• greptimedb Support Python scripting in the database
• haem A Python library for working on Bioinformatics problems.
• html-py-ever Using html5ever through kuchiki to speed up html parsing and css-selecting.
• hyperjson A hyper-fast Python module for reading/writing JSON data using Rust's serde-json.
• inline-python Inline Python code directly in your Rust code.
• johnnycanencrypt OpenPGP library with Yubikey support.
• jsonschema-rs Fast JSON Schema validation library.
• mocpy Astronomical Python library offering data structures for describing any arbitrary coverage regions on the unit sphere.
• opendal A data access layer that allows users to easily and efficiently retrieve data from various storage services in a unified way.
• orjson Fast Python JSON library.
• ormsgpack Fast Python msgpack library.
• point-process High level API for pointprocesses as a Python library.
• polaroid Hyper Fast and safe image manipulation library for Python written in Rust.
• polars Fast multi-threaded DataFrame library in Rust | Python | Node.js.
• pydantic-core Core validation logic for pydantic written in Rust.
• pyheck Fast case conversion library, built by wrapping heck.
  • Quite easy to follow as there's not much code.
• pyre Fast Python HTTP server written in Rust.
• ril-py A performant and high-level image processing library for Python written in Rust.
• river Online machine learning in python, the computationally heavy statistics algorithms are implemented in Rust.
• rust-python-coverage Example PyO3 project with automated test coverage for Rust and Python.
• tiktoken A fast BPE tokeniser for use with OpenAI's models.
• tokenizers Python bindings to the Hugging Face tokenizers (NLP) written in Rust.
• tzfpy A fast package to convert longitude/latitude to timezone name.
• utiles Fast Python web-map tile utilities
• wasmer-python Python library to run WebAssembly binaries.

Articles and other media

• (Video) Extending Python with Rust using PyO3 - Dec 16, 2023
• A Week of PyO3 + rust-numpy (How to Speed Up Your Data Pipeline X Times) - Jun 6, 2023
• (Podcast) PyO3 with David Hewitt - May 19, 2023
• Making Python 100x faster with less than 100 lines of Rust - Mar 28, 2023
• How Pydantic V2 leverages Rust's Superpowers - Feb 4, 2023
• How we extended the River stats module with Rust using PyO3 - Dec 23, 2022
• Nine Rules for Writing Python Extensions in Rust - Dec 31, 2021
• Calling Rust from Python using PyO3 - Nov 18, 2021
• davidhewitt's 2021 talk at Rust Manchester meetup - Aug 19, 2021
• Incrementally porting a small Python project to Rust - Apr 29, 2021
• Vortexa - Integrating Rust into Python - Apr 12, 2021
• Writing and publishing a Python module in Rust - Aug 2, 2020

Contributing

Everyone is welcomed to contribute to PyO3! There are many ways to support the project, such as:

• help PyO3 users with issues on GitHub and Discord
• improve documentation
• write features and bugfixes
• publish blogs and examples of how to use PyO3

Our contributing notes and architecture guide have more resources if you wish to volunteer time for PyO3 and are searching where to start.

If you don't have time to contribute yourself but still wish to support the project's future success, some of our maintainers have GitHub sponsorship pages:

• davidhewitt
• messense

License

PyO3 is licensed under the Apache-2.0 license or the MIT license, at your option.

Python is licensed under the Python License.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in PyO3 by you, as defined in the Apache License, shall be dual-licensed as above, without any additional terms or conditions.

Installation

To get started using PyO3 you will need three things: a Rust toolchain, a Python environment, and a way to build. We'll cover each of these below.

If you'd like to chat to the PyO3 maintainers and other PyO3 users, consider joining the PyO3 Discord server. We're keen to hear about your experience getting started, so we can make PyO3 as accessible as possible for everyone!

Rust

First, make sure you have Rust installed on your system. If you haven't already done so, try following the instructions here. PyO3 runs on both the stable and nightly versions so you can choose whichever one fits you best. The minimum required Rust version is 1.56.

If you can run rustc --version and the version is new enough you're good to go!

Python

To use PyO3, you need at least Python 3.7. While you can simply use the default Python interpreter on your system, it is recommended to use a virtual environment.

Virtualenvs

While you can use any virtualenv manager you like, we recommend the use of pyenv in particular if you want to develop or test for multiple different Python versions, so that is what the examples in this book will use. The installation instructions for pyenv can be found here. (Note: To get the pyenv activate and pyenv virtualenv commands, you will also need to install the pyenv-virtualenv plugin. The pyenv installer will install both together.)

If you intend to run Python from Rust (for example in unit tests) you should set the following environment variable when installing a new Python version using pyenv:

PYTHON_CONFIGURE_OPTS="--enable-shared"

For example:

env PYTHON_CONFIGURE_OPTS="--enable-shared" pyenv install 3.12

You can read more about pyenv's configuration options here.

Building

There are a number of build and Python package management systems such as setuptools-rust or manually. We recommend the use of maturin, which you can install here. It is developed to work with PyO3 and provides the most "batteries included" experience, especially if you are aiming to publish to PyPI. maturin is just a Python package, so you can add it in the same way you already install Python packages.

System Python:

pip install maturin --user

pipx:

pipx install maturin

pyenv:

pyenv activate pyo3
pip install maturin

poetry:

poetry add -G dev maturin

After installation, you can run maturin --version to check that you have correctly installed it.

Starting a new project

First you should create the folder and virtual environment that are going to contain your new project. Here we will use the recommended pyenv:

mkdir pyo3-example
cd pyo3-example
pyenv virtualenv pyo3
pyenv local pyo3

After this, you should install your build manager. In this example, we will use maturin. After you've activated your virtualenv, add maturin to it:

pip install maturin

Now you can initialize the new project:

maturin init

If maturin is already installed, you can create a new project using that directly as well:

maturin new -b pyo3 pyo3-example
cd pyo3-example
pyenv virtualenv pyo3
pyenv local pyo3

Adding to an existing project

Sadly, maturin cannot currently be run in existing projects, so if you want to use Python in an existing project you basically have two options:

1. Create a new project as above and move your existing code into that project
2. Manually edit your project configuration as necessary

If you opt for the second option, here are the things you need to pay attention to:

Cargo.toml

Make sure that the Rust crate you want to be able to access from Python is compiled into a library. You can have a binary output as well, but the code you want to access from Python has to be in the library part. Also, make sure that the crate type is cdylib and add PyO3 as a dependency as so:

# If you already have [package] information in `Cargo.toml`, you can ignore
# this section!
[package]
# `name` here is name of the package.
name = "pyo3_start"
# these are good defaults:
version = "0.1.0"
edition = "2021"

[lib]
# The name of the native library. This is the name which will be used in Python to import the
# library (i.e. `import string_sum`). If you change this, you must also change the name of the
# `#[pymodule]` in `src/lib.rs`.
name = "pyo3_example"

# "cdylib" is necessary to produce a shared library for Python to import from.
crate-type = ["cdylib"]

[dependencies]
pyo3 = { version = "0.21.2", features = ["extension-module"] }

pyproject.toml

You should also create a pyproject.toml with the following contents:

[build-system]
requires = ["maturin>=1,<2"]
build-backend = "maturin"

[project]
name = "pyo3_example"
requires-python = ">=3.7"
classifiers = [
    "Programming Language :: Rust",
    "Programming Language :: Python :: Implementation :: CPython",
    "Programming Language :: Python :: Implementation :: PyPy",
]

Running code

After this you can setup Rust code to be available in Python as below; for example, you can place this code in src/lib.rs:

use pyo3::prelude::*;

/// Formats the sum of two numbers as string.
#[pyfunction]
fn sum_as_string(a: usize, b: usize) -> PyResult<String> {
    Ok((a + b).to_string())
}

/// A Python module implemented in Rust. The name of this function must match
/// the `lib.name` setting in the `Cargo.toml`, else Python will not be able to
/// import the module.
#[pymodule]
fn pyo3_example(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(sum_as_string, m)?)?;
    Ok(())
}

Now you can run maturin develop to prepare the Python package, after which you can use it like so:

$ maturin develop
# lots of progress output as maturin runs the compilation...
$ python
>>> import pyo3_example
>>> pyo3_example.sum_as_string(5, 20)
'25'

For more instructions on how to use Python code from Rust, see the Python from Rust page.

Using Rust from Python

This chapter of the guide is dedicated to explaining how to wrap Rust code into Python objects.

PyO3 uses Rust's "procedural macros" to provide a powerful yet simple API to denote what Rust code should map into Python objects.

The three types of Python objects which PyO3 can produce are:

• Python modules, via the #[pymodule] macro
• Python functions, via the #[pyfunction] macro
• Python classes, via the #[pyclass] macro (plus #[pymethods] to define methods for those clases)

The following subchapters go through each of these in turn.

Python modules

You can create a module using #[pymodule]:

use pyo3::prelude::*;

#[pyfunction]
fn double(x: usize) -> usize {
    x * 2
}

/// This module is implemented in Rust.
#[pymodule]
fn my_extension(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(double, m)?)?;
    Ok(())
}

The #[pymodule] procedural macro takes care of exporting the initialization function of your module to Python.

The module's name defaults to the name of the Rust function. You can override the module name by using #[pyo3(name = "custom_name")]:

use pyo3::prelude::*;

#[pyfunction]
fn double(x: usize) -> usize {
    x * 2
}

#[pymodule]
#[pyo3(name = "custom_name")]
fn my_extension(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(double, m)?)?;
    Ok(())
}

The name of the module must match the name of the .so or .pyd file. Otherwise, you will get an import error in Python with the following message: ImportError: dynamic module does not define module export function (PyInit_name_of_your_module)

To import the module, either:

• copy the shared library as described in Manual builds, or
• use a tool, e.g. maturin develop with maturin or python setup.py develop with setuptools-rust.

Documentation

The Rust doc comments of the module initialization function will be applied automatically as the Python docstring of your module.

For example, building off of the above code, this will print This module is implemented in Rust.:

import my_extension

print(my_extension.__doc__)

Python submodules

You can create a module hierarchy within a single extension module by using Bound<'_, PyModule>::add_submodule(). For example, you could define the modules parent_module and parent_module.child_module.

use pyo3::prelude::*;

#[pymodule]
fn parent_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    register_child_module(m)?;
    Ok(())
}

fn register_child_module(parent_module: &Bound<'_, PyModule>) -> PyResult<()> {
    let child_module = PyModule::new_bound(parent_module.py(), "child_module")?;
    child_module.add_function(wrap_pyfunction!(func, &child_module)?)?;
    parent_module.add_submodule(&child_module)?;
    Ok(())
}

#[pyfunction]
fn func() -> String {
    "func".to_string()
}

Python::with_gil(|py| {
   use pyo3::wrap_pymodule;
   use pyo3::types::IntoPyDict;
   let parent_module = wrap_pymodule!(parent_module)(py);
   let ctx = [("parent_module", parent_module)].into_py_dict_bound(py);

   py.run_bound("assert parent_module.child_module.func() == 'func'", None, Some(&ctx)).unwrap();
})

Note that this does not define a package, so this won’t allow Python code to directly import submodules by using from parent_module import child_module. For more information, see #759 and #1517.

It is not necessary to add #[pymodule] on nested modules, which is only required on the top-level module.

Declarative modules (experimental)

Another syntax based on Rust inline modules is also available to declare modules. The experimental-declarative-modules feature must be enabled to use it.

For example:

#[cfg(feature = "experimental-declarative-modules")]
mod declarative_module_test {
use pyo3::prelude::*;

#[pyfunction]
fn double(x: usize) -> usize {
    x * 2
}

#[pymodule]
mod my_extension {
    use super::*;

    #[pymodule_export]
    use super::double; // Exports the double function as part of the module

    #[pyfunction] // This will be part of the module
    fn triple(x: usize) -> usize {
        x * 3
    }

    #[pyclass] // This will be part of the module
    struct Unit;

    #[pymodule]
    mod submodule {
        // This is a submodule
    }

    #[pymodule_init]
    fn init(m: &Bound<'_, PyModule>) -> PyResult<()> {
        // Arbitrary code to run at the module initialization
        m.add("double2", m.getattr("double")?)?;
        Ok(())
    }
}
}

Some changes are planned to this feature before stabilization, like automatically filling submodules into sys.modules to allow easier imports (see issue #759) and filling the module argument of inlined #[pyclass] automatically with the proper module name. Macro names might also change. See issue #3900 to track this feature progress.

Python functions

The #[pyfunction] attribute is used to define a Python function from a Rust function. Once defined, the function needs to be added to a module using the wrap_pyfunction! macro.

The following example defines a function called double in a Python module called my_extension:

use pyo3::prelude::*;

#[pyfunction]
fn double(x: usize) -> usize {
    x * 2
}

#[pymodule]
fn my_extension(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(double, m)?)?;
    Ok(())
}

This chapter of the guide explains full usage of the #[pyfunction] attribute. In this first section, the following topics are covered:

• Function options
  • #[pyo3(name = "...")]
  • #[pyo3(signature = (...))]
  • #[pyo3(text_signature = "...")]
  • #[pyo3(pass_module)]
• Per-argument options
• Advanced function patterns
• #[pyfn] shorthand

There are also additional sections on the following topics:

• Function Signatures

Function options

The #[pyo3] attribute can be used to modify properties of the generated Python function. It can take any combination of the following options:

• #[pyo3(name = "...")]

  Overrides the name exposed to Python.

  In the following example, the Rust function no_args_py will be added to the Python module module_with_functions as the Python function no_args:

  use pyo3::prelude::*;
  
  #[pyfunction]
  #[pyo3(name = "no_args")]
  fn no_args_py() -> usize {
      42
  }
  
  #[pymodule]
  fn module_with_functions(m: &Bound<'_, PyModule>) -> PyResult<()> {
      m.add_function(wrap_pyfunction!(no_args_py, m)?)?;
      Ok(())
  }
  
  Python::with_gil(|py| {
      let m = pyo3::wrap_pymodule!(module_with_functions)(py);
      assert!(m.getattr(py, "no_args").is_ok());
      assert!(m.getattr(py, "no_args_py").is_err());
  });

• #[pyo3(signature = (...))]

  Defines the function signature in Python. See Function Signatures.

• #[pyo3(text_signature = "...")]

  Overrides the PyO3-generated function signature visible in Python tooling (such as via inspect.signature). See the corresponding topic in the Function Signatures subchapter.

• #[pyo3(pass_module)]

  Set this option to make PyO3 pass the containing module as the first argument to the function. It is then possible to use the module in the function body. The first argument must be of type &Bound<'_, PyModule>, Bound<'_, PyModule>, or Py<PyModule>.

  The following example creates a function pyfunction_with_module which returns the containing module's name (i.e. module_with_fn):

  use pyo3::prelude::*;
  use pyo3::types::PyString;
  
  #[pyfunction]
  #[pyo3(pass_module)]
  fn pyfunction_with_module<'py>(
      module: &Bound<'py, PyModule>,
  ) -> PyResult<Bound<'py, PyString>> {
      module.name()
  }
  
  #[pymodule]
  fn module_with_fn(m: &Bound<'_, PyModule>) -> PyResult<()> {
      m.add_function(wrap_pyfunction!(pyfunction_with_module, m)?)
  }

Per-argument options

The #[pyo3] attribute can be used on individual arguments to modify properties of them in the generated function. It can take any combination of the following options:

• #[pyo3(from_py_with = "...")]

  Set this on an option to specify a custom function to convert the function argument from Python to the desired Rust type, instead of using the default FromPyObject extraction. The function signature must be fn(&Bound<'_, PyAny>) -> PyResult<T> where T is the Rust type of the argument.

  The following example uses from_py_with to convert the input Python object to its length:

  use pyo3::prelude::*;
  
  fn get_length(obj: &Bound<'_, PyAny>) -> PyResult<usize> {
      let length = obj.len()?;
      Ok(length)
  }
  
  #[pyfunction]
  fn object_length(#[pyo3(from_py_with = "get_length")] argument: usize) -> usize {
      argument
  }
  
  Python::with_gil(|py| {
      let f = pyo3::wrap_pyfunction_bound!(object_length)(py).unwrap();
      assert_eq!(f.call1((vec![1, 2, 3],)).unwrap().extract::<usize>().unwrap(), 3);
  });

Advanced function patterns

Calling Python functions in Rust

You can pass Python def'd functions and built-in functions to Rust functions PyFunction corresponds to regular Python functions while PyCFunction describes built-ins such as repr().

You can also use Bound<'_, PyAny>::is_callable to check if you have a callable object. is_callable will return true for functions (including lambdas), methods and objects with a __call__ method. You can call the object with Bound<'_, PyAny>::call with the args as first parameter and the kwargs (or None) as second parameter. There are also Bound<'_, PyAny>::call0 with no args and Bound<'_, PyAny>::call1 with only positional args.

Calling Rust functions in Python

The ways to convert a Rust function into a Python object vary depending on the function:

• Named functions, e.g. fn foo(): add #[pyfunction] and then use wrap_pyfunction! to get the corresponding PyCFunction.
• Anonymous functions (or closures), e.g. foo: fn() either:
  • use a #[pyclass] struct which stores the function as a field and implement __call__ to call the stored function.
  • use PyCFunction::new_closure to create an object directly from the function.

Accessing the FFI functions

In order to make Rust functions callable from Python, PyO3 generates an extern "C" function whose exact signature depends on the Rust signature. (PyO3 chooses the optimal Python argument passing convention.) It then embeds the call to the Rust function inside this FFI-wrapper function. This wrapper handles extraction of the regular arguments and the keyword arguments from the input PyObjects.

The wrap_pyfunction macro can be used to directly get a Bound<PyCFunction> given a #[pyfunction] and a Bound<PyModule>: wrap_pyfunction!(rust_fun, module).

#[pyfn] shorthand

There is a shorthand to #[pyfunction] and wrap_pymodule!: the function can be placed inside the module definition and annotated with #[pyfn]. To simplify PyO3, it is expected that #[pyfn] may be removed in a future release (See #694).

An example of #[pyfn] is below:

use pyo3::prelude::*;

#[pymodule]
fn my_extension(m: &Bound<'_, PyModule>) -> PyResult<()> {
    #[pyfn(m)]
    fn double(x: usize) -> usize {
        x * 2
    }

    Ok(())
}

#[pyfn(m)] is just syntactic sugar for #[pyfunction], and takes all the same options documented in the rest of this chapter. The code above is expanded to the following:

use pyo3::prelude::*;

#[pymodule]
fn my_extension(m: &Bound<'_, PyModule>) -> PyResult<()> {
    #[pyfunction]
    fn double(x: usize) -> usize {
        x * 2
    }

    m.add_function(wrap_pyfunction!(double, m)?)?;
    Ok(())
}

Function signatures

The #[pyfunction] attribute also accepts parameters to control how the generated Python function accepts arguments. Just like in Python, arguments can be positional-only, keyword-only, or accept either. *args lists and **kwargs dicts can also be accepted. These parameters also work for #[pymethods] which will be introduced in the Python Classes section of the guide.

Like Python, by default PyO3 accepts all arguments as either positional or keyword arguments. Most arguments are required by default, except for trailing Option<_> arguments, which are implicitly given a default of None. This behaviour can be configured by the #[pyo3(signature = (...))] option which allows writing a signature in Python syntax.

This section of the guide goes into detail about use of the #[pyo3(signature = (...))] option and its related option #[pyo3(text_signature = "...")]

Using #[pyo3(signature = (...))]

For example, below is a function that accepts arbitrary keyword arguments (**kwargs in Python syntax) and returns the number that was passed:

use pyo3::prelude::*;
use pyo3::types::PyDict;

#[pyfunction]
#[pyo3(signature = (**kwds))]
fn num_kwds(kwds: Option<&Bound<'_, PyDict>>) -> usize {
    kwds.map_or(0, |dict| dict.len())
}

#[pymodule]
fn module_with_functions(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(num_kwds, m)?).unwrap();
    Ok(())
}

Just like in Python, the following constructs can be part of the signature::

• /: positional-only arguments separator, each parameter defined before / is a positional-only parameter.
• *: var arguments separator, each parameter defined after * is a keyword-only parameter.
• *args: "args" is var args. Type of the args parameter has to be &Bound<'_, PyTuple>.
• **kwargs: "kwargs" receives keyword arguments. The type of the kwargs parameter has to be Option<&Bound<'_, PyDict>>.
• arg=Value: arguments with default value. If the arg argument is defined after var arguments, it is treated as a keyword-only argument. Note that Value has to be valid rust code, PyO3 just inserts it into the generated code unmodified.

Example:

use pyo3::prelude::*;
use pyo3::types::{PyDict, PyTuple};

#[pyclass]
struct MyClass {
    num: i32,
}
#[pymethods]
impl MyClass {
    #[new]
    #[pyo3(signature = (num=-1))]
    fn new(num: i32) -> Self {
        MyClass { num }
    }

    #[pyo3(signature = (num=10, *py_args, name="Hello", **py_kwargs))]
    fn method(
        &mut self,
        num: i32,
        py_args: &Bound<'_, PyTuple>,
        name: &str,
        py_kwargs: Option<&Bound<'_, PyDict>>,
    ) -> String {
        let num_before = self.num;
        self.num = num;
        format!(
            "num={} (was previously={}), py_args={:?}, name={}, py_kwargs={:?} ",
            num, num_before, py_args, name, py_kwargs,
        )
    }

    fn make_change(&mut self, num: i32) -> PyResult<String> {
        self.num = num;
        Ok(format!("num={}", self.num))
    }
}

Arguments of type Python must not be part of the signature:

#![allow(dead_code)]
use pyo3::prelude::*;
#[pyfunction]
#[pyo3(signature = (lambda))]
pub fn simple_python_bound_function(py: Python<'_>, lambda: PyObject) -> PyResult<()> {
    Ok(())
}

N.B. the position of the / and * arguments (if included) control the system of handling positional and keyword arguments. In Python:

import mymodule

mc = mymodule.MyClass()
print(mc.method(44, False, "World", 666, x=44, y=55))
print(mc.method(num=-1, name="World"))
print(mc.make_change(44, False))

Produces output:

py_args=('World', 666), py_kwargs=Some({'x': 44, 'y': 55}), name=Hello, num=44
py_args=(), py_kwargs=None, name=World, num=-1
num=44
num=-1

Note: to use keywords like struct as a function argument, use "raw identifier" syntax r#struct in both the signature and the function definition:

#![allow(dead_code)]
use pyo3::prelude::*;
#[pyfunction(signature = (r#struct = "foo"))]
fn function_with_keyword(r#struct: &str) {
    let _ = r#struct;
    /* ... */
}

Trailing optional arguments

As a convenience, functions without a #[pyo3(signature = (...))] option will treat trailing Option<T> arguments as having a default of None. In the example below, PyO3 will create increment with a signature of increment(x, amount=None).

use pyo3::prelude::*;

/// Returns a copy of `x` increased by `amount`.
///
/// If `amount` is unspecified or `None`, equivalent to `x + 1`.
#[pyfunction]
fn increment(x: u64, amount: Option<u64>) -> u64 {
    x + amount.unwrap_or(1)
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let fun = pyo3::wrap_pyfunction_bound!(increment, py)?;

        let inspect = PyModule::import_bound(py, "inspect")?.getattr("signature")?;
        let sig: String = inspect
            .call1((fun,))?
            .call_method0("__str__")?
            .extract()?;

        #[cfg(Py_3_8)]  // on 3.7 the signature doesn't render b, upstream bug?
        assert_eq!(sig, "(x, amount=None)");

        Ok(())
    })
}

To make trailing Option<T> arguments required, but still accept None, add a #[pyo3(signature = (...))] annotation. For the example above, this would be #[pyo3(signature = (x, amount))]:

use pyo3::prelude::*;
#[pyfunction]
#[pyo3(signature = (x, amount))]
fn increment(x: u64, amount: Option<u64>) -> u64 {
    x + amount.unwrap_or(1)
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let fun = pyo3::wrap_pyfunction_bound!(increment, py)?;

        let inspect = PyModule::import_bound(py, "inspect")?.getattr("signature")?;
        let sig: String = inspect
            .call1((fun,))?
            .call_method0("__str__")?
            .extract()?;

        #[cfg(Py_3_8)]  // on 3.7 the signature doesn't render b, upstream bug?
        assert_eq!(sig, "(x, amount)");

        Ok(())
    })
}

To help avoid confusion, PyO3 requires #[pyo3(signature = (...))] when an Option<T> argument is surrounded by arguments which aren't Option<T>.

Making the function signature available to Python

The function signature is exposed to Python via the __text_signature__ attribute. PyO3 automatically generates this for every #[pyfunction] and all #[pymethods] directly from the Rust function, taking into account any override done with the #[pyo3(signature = (...))] option.

This automatic generation can only display the value of default arguments for strings, integers, boolean types, and None. Any other default arguments will be displayed as .... (.pyi type stub files commonly also use ... for default arguments in the same way.)

In cases where the automatically-generated signature needs adjusting, it can be overridden using the #[pyo3(text_signature)] option.)

The example below creates a function add which accepts two positional-only arguments a and b, where b has a default value of zero.

use pyo3::prelude::*;

/// This function adds two unsigned 64-bit integers.
#[pyfunction]
#[pyo3(signature = (a, b=0, /))]
fn add(a: u64, b: u64) -> u64 {
    a + b
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let fun = pyo3::wrap_pyfunction_bound!(add, py)?;

        let doc: String = fun.getattr("__doc__")?.extract()?;
        assert_eq!(doc, "This function adds two unsigned 64-bit integers.");

        let inspect = PyModule::import_bound(py, "inspect")?.getattr("signature")?;
        let sig: String = inspect
            .call1((fun,))?
            .call_method0("__str__")?
            .extract()?;

        #[cfg(Py_3_8)]  // on 3.7 the signature doesn't render b, upstream bug?
        assert_eq!(sig, "(a, b=0, /)");

        Ok(())
    })
}

The following IPython output demonstrates how this generated signature will be seen from Python tooling:

>>> pyo3_test.add.__text_signature__
'(a, b=..., /)'
>>> pyo3_test.add?
Signature: pyo3_test.add(a, b=0, /)
Docstring: This function adds two unsigned 64-bit integers.
Type:      builtin_function_or_method

Overriding the generated signature

The #[pyo3(text_signature = "(<some signature>)")] attribute can be used to override the default generated signature.

In the snippet below, the text signature attribute is used to include the default value of 0 for the argument b, instead of the automatically-generated default value of ...:

use pyo3::prelude::*;

/// This function adds two unsigned 64-bit integers.
#[pyfunction]
#[pyo3(signature = (a, b=0, /), text_signature = "(a, b=0, /)")]
fn add(a: u64, b: u64) -> u64 {
    a + b
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let fun = pyo3::wrap_pyfunction_bound!(add, py)?;

        let doc: String = fun.getattr("__doc__")?.extract()?;
        assert_eq!(doc, "This function adds two unsigned 64-bit integers.");

        let inspect = PyModule::import_bound(py, "inspect")?.getattr("signature")?;
        let sig: String = inspect
            .call1((fun,))?
            .call_method0("__str__")?
            .extract()?;
        assert_eq!(sig, "(a, b=0, /)");

        Ok(())
    })
}

PyO3 will include the contents of the annotation unmodified as the __text_signature__. Below shows how IPython will now present this (see the default value of 0 for b):

>>> pyo3_test.add.__text_signature__
'(a, b=0, /)'
>>> pyo3_test.add?
Signature: pyo3_test.add(a, b=0, /)
Docstring: This function adds two unsigned 64-bit integers.
Type:      builtin_function_or_method

If no signature is wanted at all, #[pyo3(text_signature = None)] will disable the built-in signature. The snippet below demonstrates use of this:

use pyo3::prelude::*;

/// This function adds two unsigned 64-bit integers.
#[pyfunction]
#[pyo3(signature = (a, b=0, /), text_signature = None)]
fn add(a: u64, b: u64) -> u64 {
    a + b
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let fun = pyo3::wrap_pyfunction_bound!(add, py)?;

        let doc: String = fun.getattr("__doc__")?.extract()?;
        assert_eq!(doc, "This function adds two unsigned 64-bit integers.");
        assert!(fun.getattr("__text_signature__")?.is_none());

        Ok(())
    })
}

Now the function's __text_signature__ will be set to None, and IPython will not display any signature in the help:

>>> pyo3_test.add.__text_signature__ == None
True
>>> pyo3_test.add?
Docstring: This function adds two unsigned 64-bit integers.
Type:      builtin_function_or_method

Error handling

This chapter contains a little background of error handling in Rust and how PyO3 integrates this with Python exceptions.

This covers enough detail to create a #[pyfunction] which raises Python exceptions from errors originating in Rust.

There is a later section of the guide on Python exceptions which covers exception types in more detail.

Representing Python exceptions

Rust code uses the generic Result<T, E> enum to propagate errors. The error type E is chosen by the code author to describe the possible errors which can happen.

PyO3 has the PyErr type which represents a Python exception. If a PyO3 API could result in a Python exception being raised, the return type of that API will be PyResult<T>, which is an alias for the type Result<T, PyErr>.

In summary:

• When Python exceptions are raised and caught by PyO3, the exception will be stored in the Err variant of the PyResult.
• Passing Python exceptions through Rust code then uses all the "normal" techniques such as the ? operator, with PyErr as the error type.
• Finally, when a PyResult crosses from Rust back to Python via PyO3, if the result is an Err variant the contained exception will be raised.

(There are many great tutorials on Rust error handling and the ? operator, so this guide will not go into detail on Rust-specific topics.)

Raising an exception from a function

As indicated in the previous section, when a PyResult containing an Err crosses from Rust to Python, PyO3 will raise the exception contained within.

Accordingly, to raise an exception from a #[pyfunction], change the return type T to PyResult<T>. When the function returns an Err it will raise a Python exception. (Other Result<T, E> types can be used as long as the error E has a From conversion for PyErr, see custom Rust error types below.)

This also works for functions in #[pymethods].

For example, the following check_positive function raises a ValueError when the input is negative:

use pyo3::exceptions::PyValueError;
use pyo3::prelude::*;

#[pyfunction]
fn check_positive(x: i32) -> PyResult<()> {
    if x < 0 {
        Err(PyValueError::new_err("x is negative"))
    } else {
        Ok(())
    }
}

fn main(){
	Python::with_gil(|py|{
		let fun = pyo3::wrap_pyfunction_bound!(check_positive, py).unwrap();
		fun.call1((-1,)).unwrap_err();
		fun.call1((1,)).unwrap();
	});
}

All built-in Python exception types are defined in the pyo3::exceptions module. They have a new_err constructor to directly build a PyErr, as seen in the example above.

Custom Rust error types

PyO3 will automatically convert a Result<T, E> returned by a #[pyfunction] into a PyResult<T> as long as there is an implementation of std::from::From<E> for PyErr. Many error types in the Rust standard library have a From conversion defined in this way.

If the type E you are handling is defined in a third-party crate, see the section on foreign rust error types below for ways to work with this error.

The following example makes use of the implementation of From<ParseIntError> for PyErr to raise exceptions encountered when parsing strings as integers:

use pyo3::prelude::*;
use std::num::ParseIntError;

#[pyfunction]
fn parse_int(x: &str) -> Result<usize, ParseIntError> {
    x.parse()
}

fn main() {
    Python::with_gil(|py| {
        let fun = pyo3::wrap_pyfunction_bound!(parse_int, py).unwrap();
        let value: usize = fun.call1(("5",)).unwrap().extract().unwrap();
        assert_eq!(value, 5);
    });
}

When passed a string which doesn't contain a floating-point number, the exception raised will look like the below:

>>> parse_int("bar")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: invalid digit found in string

As a more complete example, the following snippet defines a Rust error named CustomIOError. It then defines a From<CustomIOError> for PyErr, which returns a PyErr representing Python's OSError. Therefore, it can use this error in the result of a #[pyfunction] directly, relying on the conversion if it has to be propagated into a Python exception.

use pyo3::exceptions::PyOSError;
use pyo3::prelude::*;
use std::fmt;

#[derive(Debug)]
struct CustomIOError;

impl std::error::Error for CustomIOError {}

impl fmt::Display for CustomIOError {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(f, "Oh no!")
    }
}

impl std::convert::From<CustomIOError> for PyErr {
    fn from(err: CustomIOError) -> PyErr {
        PyOSError::new_err(err.to_string())
    }
}

pub struct Connection {/* ... */}

fn bind(addr: String) -> Result<Connection, CustomIOError> {
    if &addr == "0.0.0.0" {
        Err(CustomIOError)
    } else {
        Ok(Connection{ /* ... */})
    }
}

#[pyfunction]
fn connect(s: String) -> Result<(), CustomIOError> {
    bind(s)?;
    // etc.
    Ok(())
}

fn main() {
    Python::with_gil(|py| {
        let fun = pyo3::wrap_pyfunction_bound!(connect, py).unwrap();
        let err = fun.call1(("0.0.0.0",)).unwrap_err();
        assert!(err.is_instance_of::<PyOSError>(py));
    });
}

If lazy construction of the Python exception instance is desired, the PyErrArguments trait can be implemented instead of From. In that case, actual exception argument creation is delayed until the PyErr is needed.

A final note is that any errors E which have a From conversion can be used with the ? ("try") operator with them. An alternative implementation of the above parse_int which instead returns PyResult is below:

use pyo3::prelude::*;

fn parse_int(s: String) -> PyResult<usize> {
    let x = s.parse()?;
    Ok(x)
}

use pyo3::exceptions::PyValueError;

fn main() {
    Python::with_gil(|py| {
        assert_eq!(parse_int(String::from("1")).unwrap(), 1);
        assert_eq!(parse_int(String::from("1337")).unwrap(), 1337);

        assert!(parse_int(String::from("-1"))
            .unwrap_err()
            .is_instance_of::<PyValueError>(py));
        assert!(parse_int(String::from("foo"))
            .unwrap_err()
            .is_instance_of::<PyValueError>(py));
        assert!(parse_int(String::from("13.37"))
            .unwrap_err()
            .is_instance_of::<PyValueError>(py));
    })
}

Foreign Rust error types

The Rust compiler will not permit implementation of traits for types outside of the crate where the type is defined. (This is known as the "orphan rule".)

Given a type OtherError which is defined in third-party code, there are two main strategies available to integrate it with PyO3:

• Create a newtype wrapper, e.g. MyOtherError. Then implement From<MyOtherError> for PyErr (or PyErrArguments), as well as From<OtherError> for MyOtherError.
• Use Rust's Result combinators such as map_err to write code freely to convert OtherError into whatever is needed. This requires boilerplate at every usage however gives unlimited flexibility.

To detail the newtype strategy a little further, the key trick is to return Result<T, MyOtherError> from the #[pyfunction]. This means that PyO3 will make use of From<MyOtherError> for PyErr to create Python exceptions while the #[pyfunction] implementation can use ? to convert OtherError to MyOtherError automatically.

The following example demonstrates this for some imaginary third-party crate some_crate with a function get_x returning Result<i32, OtherError>:

mod some_crate {
  pub struct OtherError(());
  impl OtherError {
      pub fn message(&self) -> &'static str { "some error occurred" }
  }
  pub fn get_x() -> Result<i32, OtherError> { Ok(5) }
}

use pyo3::prelude::*;
use pyo3::exceptions::PyValueError;
use some_crate::{OtherError, get_x};

struct MyOtherError(OtherError);

impl From<MyOtherError> for PyErr {
    fn from(error: MyOtherError) -> Self {
        PyValueError::new_err(error.0.message())
    }
}

impl From<OtherError> for MyOtherError {
    fn from(other: OtherError) -> Self {
        Self(other)
    }
}

#[pyfunction]
fn wrapped_get_x() -> Result<i32, MyOtherError> {
    // get_x is a function returning Result<i32, OtherError>
    let x: i32 = get_x()?;
    Ok(x)
}

fn main() {
    Python::with_gil(|py| {
        let fun = pyo3::wrap_pyfunction_bound!(wrapped_get_x, py).unwrap();
        let value: usize = fun.call0().unwrap().extract().unwrap();
        assert_eq!(value, 5);
    });
}

Python classes

PyO3 exposes a group of attributes powered by Rust's proc macro system for defining Python classes as Rust structs.

The main attribute is #[pyclass], which is placed upon a Rust struct or enum to generate a Python type for it. They will usually also have one #[pymethods]-annotated impl block for the struct, which is used to define Python methods and constants for the generated Python type. (If the multiple-pymethods feature is enabled, each #[pyclass] is allowed to have multiple #[pymethods] blocks.) #[pymethods] may also have implementations for Python magic methods such as __str__.

This chapter will discuss the functionality and configuration these attributes offer. Below is a list of links to the relevant section of this chapter for each:

• #[pyclass]
  • #[pyo3(get, set)]
• #[pymethods]
  • #[new]
  • #[getter]
  • #[setter]
  • #[staticmethod]
  • #[classmethod]
  • #[classattr]
  • #[args]
• Magic methods and slots
• Classes as function arguments

Defining a new class

To define a custom Python class, add the #[pyclass] attribute to a Rust struct or enum.

#![allow(dead_code)]
use pyo3::prelude::*;

#[pyclass]
struct MyClass {
    inner: i32,
}

// A "tuple" struct
#[pyclass]
struct Number(i32);

// PyO3 supports unit-only enums (which contain only unit variants)
// These simple enums behave similarly to Python's enumerations (enum.Enum)
#[pyclass]
enum MyEnum {
    Variant,
    OtherVariant = 30, // PyO3 supports custom discriminants.
}

// PyO3 supports custom discriminants in unit-only enums
#[pyclass]
enum HttpResponse {
    Ok = 200,
    NotFound = 404,
    Teapot = 418,
    // ...
}

// PyO3 also supports enums with non-unit variants
// These complex enums have sligtly different behavior from the simple enums above
// They are meant to work with instance checks and match statement patterns
#[pyclass]
enum Shape {
    Circle { radius: f64 },
    Rectangle { width: f64, height: f64 },
    RegularPolygon { side_count: u32, radius: f64 },
    Nothing {},
}

The above example generates implementations for PyTypeInfo and PyClass for MyClass, Number, MyEnum, HttpResponse, and Shape. To see these generated implementations, refer to the implementation details at the end of this chapter.

Restrictions

To integrate Rust types with Python, PyO3 needs to place some restrictions on the types which can be annotated with #[pyclass]. In particular, they must have no lifetime parameters, no generic parameters, and must implement Send. The reason for each of these is explained below.

No lifetime parameters

Rust lifetimes are used by the Rust compiler to reason about a program's memory safety. They are a compile-time only concept; there is no way to access Rust lifetimes at runtime from a dynamic language like Python.

As soon as Rust data is exposed to Python, there is no guarantee that the Rust compiler can make on how long the data will live. Python is a reference-counted language and those references can be held for an arbitrarily long time which is untraceable by the Rust compiler. The only possible way to express this correctly is to require that any #[pyclass] does not borrow data for any lifetime shorter than the 'static lifetime, i.e. the #[pyclass] cannot have any lifetime parameters.

When you need to share ownership of data between Python and Rust, instead of using borrowed references with lifetimes consider using reference-counted smart pointers such as Arc or Py.

No generic parameters

A Rust struct Foo<T> with a generic parameter T generates new compiled implementations each time it is used with a different concrete type for T. These new implementations are generated by the compiler at each usage site. This is incompatible with wrapping Foo in Python, where there needs to be a single compiled implementation of Foo which is integrated with the Python interpreter.

Currently, the best alternative is to write a macro which expands to a new #[pyclass] for each instantiation you want:

#![allow(dead_code)]
use pyo3::prelude::*;

struct GenericClass<T> {
    data: T,
}

macro_rules! create_interface {
    ($name: ident, $type: ident) => {
        #[pyclass]
        pub struct $name {
            inner: GenericClass<$type>,
        }
        #[pymethods]
        impl $name {
            #[new]
            pub fn new(data: $type) -> Self {
                Self {
                    inner: GenericClass { data: data },
                }
            }
        }
    };
}

create_interface!(IntClass, i64);
create_interface!(FloatClass, String);

Must be Send

Because Python objects are freely shared between threads by the Python interpreter, there is no guarantee which thread will eventually drop the object. Therefore all types annotated with #[pyclass] must implement Send (unless annotated with #[pyclass(unsendable)]).

Constructor

By default, it is not possible to create an instance of a custom class from Python code. To declare a constructor, you need to define a method and annotate it with the #[new] attribute. Only Python's __new__ method can be specified, __init__ is not available.

#![allow(dead_code)]
use pyo3::prelude::*;
#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    #[new]
    fn new(value: i32) -> Self {
        Number(value)
    }
}

Alternatively, if your new method may fail you can return PyResult<Self>.

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::exceptions::PyValueError;
#[pyclass]
struct Nonzero(i32);

#[pymethods]
impl Nonzero {
    #[new]
    fn py_new(value: i32) -> PyResult<Self> {
        if value == 0 {
            Err(PyValueError::new_err("cannot be zero"))
        } else {
            Ok(Nonzero(value))
        }
    }
}

If you want to return an existing object (for example, because your new method caches the values it returns), new can return pyo3::Py<Self>.

As you can see, the Rust method name is not important here; this way you can still, use new() for a Rust-level constructor.

If no method marked with #[new] is declared, object instances can only be created from Rust, but not from Python.

For arguments, see the Method arguments section below.

Adding the class to a module

The next step is to create the module initializer and add our class to it:

#![allow(dead_code)]
use pyo3::prelude::*;
#[pyclass]
struct Number(i32);

#[pymodule]
fn my_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_class::<Number>()?;
    Ok(())
}

Bound and interior mutability

Often is useful to turn a #[pyclass] type T into a Python object and access it from Rust code. The [Py<T>] and [Bound<'py, T>] smart pointers are the ways to represent a Python object in PyO3's API. More detail can be found about them in the Python objects section of the guide.

Most Python objects do not offer exclusive (&mut) access (see the section on Python's memory model). However, Rust structs wrapped as Python objects (called pyclass types) often do need &mut access. Due to the GIL, PyO3 can guarantee exclusive access to them.

The Rust borrow checker cannot reason about &mut references once an object's ownership has been passed to the Python interpreter. This means that borrow checking is done at runtime using with a scheme very similar to std::cell::RefCell<T>. This is known as interior mutability.

Users who are familiar with RefCell<T> can use Py<T> and Bound<'py, T> just like RefCell<T>.

For users who are not very familiar with RefCell<T>, here is a reminder of Rust's rules of borrowing:

• At any given time, you can have either (but not both of) one mutable reference or any number of immutable references.
• References can never outlast the data they refer to.

Py<T> and Bound<'py, T>, like RefCell<T>, ensure these borrowing rules by tracking references at runtime.

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    #[pyo3(get)]
    num: i32,
}
Python::with_gil(|py| {
    let obj = Bound::new(py, MyClass { num: 3 }).unwrap();
    {
        let obj_ref = obj.borrow(); // Get PyRef
        assert_eq!(obj_ref.num, 3);
        // You cannot get PyRefMut unless all PyRefs are dropped
        assert!(obj.try_borrow_mut().is_err());
    }
    {
        let mut obj_mut = obj.borrow_mut(); // Get PyRefMut
        obj_mut.num = 5;
        // You cannot get any other refs until the PyRefMut is dropped
        assert!(obj.try_borrow().is_err());
        assert!(obj.try_borrow_mut().is_err());
    }

    // You can convert `Bound` to a Python object
    pyo3::py_run!(py, obj, "assert obj.num == 5");
});

A Bound<'py, T> is restricted to the GIL lifetime 'py. To make the object longer lived (for example, to store it in a struct on the Rust side), use Py<T>. Py<T> needs a Python<'_> token to allow access:

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    num: i32,
}

fn return_myclass() -> Py<MyClass> {
    Python::with_gil(|py| Py::new(py, MyClass { num: 1 }).unwrap())
}

let obj = return_myclass();

Python::with_gil(|py| {
    let bound = obj.bind(py); // Py<MyClass>::bind returns &Bound<'py, MyClass>
    let obj_ref = bound.borrow(); // Get PyRef<T>
    assert_eq!(obj_ref.num, 1);
});

frozen classes: Opting out of interior mutability

As detailed above, runtime borrow checking is currently enabled by default. But a class can opt of out it by declaring itself frozen. It can still use interior mutability via standard Rust types like RefCell or Mutex, but it is not bound to the implementation provided by PyO3 and can choose the most appropriate strategy on field-by-field basis.

Classes which are frozen and also Sync, e.g. they do use Mutex but not RefCell, can be accessed without needing the Python GIL via the Bound::get and Py::get methods:

use std::sync::atomic::{AtomicUsize, Ordering};
use pyo3::prelude::*;

#[pyclass(frozen)]
struct FrozenCounter {
    value: AtomicUsize,
}

let py_counter: Py<FrozenCounter> = Python::with_gil(|py| {
    let counter = FrozenCounter {
        value: AtomicUsize::new(0),
    };

    Py::new(py, counter).unwrap()
});

py_counter.get().value.fetch_add(1, Ordering::Relaxed);

Frozen classes are likely to become the default thereby guiding the PyO3 ecosystem towards a more deliberate application of interior mutability. Eventually, this should enable further optimizations of PyO3's internals and avoid downstream code paying the cost of interior mutability when it is not actually required.

Customizing the class

#[pyclass] can be used with the following parameters:

All of these parameters can either be passed directly on the #[pyclass(...)] annotation, or as one or more accompanying #[pyo3(...)] annotations, e.g.:

// Argument supplied directly to the `#[pyclass]` annotation.
#[pyclass(name = "SomeName", subclass)]
struct MyClass {}

// Argument supplied as a separate annotation.
#[pyclass]
#[pyo3(name = "SomeName", subclass)]
struct MyClass {}

These parameters are covered in various sections of this guide.

Return type

Generally, #[new] methods have to return T: Into<PyClassInitializer<Self>> or PyResult<T> where T: Into<PyClassInitializer<Self>>.

For constructors that may fail, you should wrap the return type in a PyResult as well. Consult the table below to determine which type your constructor should return:

Inheritance

By default, object, i.e. PyAny is used as the base class. To override this default, use the extends parameter for pyclass with the full path to the base class. Currently, only classes defined in Rust and builtins provided by PyO3 can be inherited from; inheriting from other classes defined in Python is not yet supported (#991).

For convenience, (T, U) implements Into<PyClassInitializer<T>> where U is the base class of T. But for a more deeply nested inheritance, you have to return PyClassInitializer<T> explicitly.

To get a parent class from a child, use PyRef instead of &self for methods, or PyRefMut instead of &mut self. Then you can access a parent class by self_.as_ref() as &Self::BaseClass, or by self_.into_super() as PyRef<Self::BaseClass>.

use pyo3::prelude::*;

#[pyclass(subclass)]
struct BaseClass {
    val1: usize,
}

#[pymethods]
impl BaseClass {
    #[new]
    fn new() -> Self {
        BaseClass { val1: 10 }
    }

    pub fn method(&self) -> PyResult<usize> {
        Ok(self.val1)
    }
}

#[pyclass(extends=BaseClass, subclass)]
struct SubClass {
    val2: usize,
}

#[pymethods]
impl SubClass {
    #[new]
    fn new() -> (Self, BaseClass) {
        (SubClass { val2: 15 }, BaseClass::new())
    }

    fn method2(self_: PyRef<'_, Self>) -> PyResult<usize> {
        let super_ = self_.as_ref(); // Get &BaseClass
        super_.method().map(|x| x * self_.val2)
    }
}

#[pyclass(extends=SubClass)]
struct SubSubClass {
    val3: usize,
}

#[pymethods]
impl SubSubClass {
    #[new]
    fn new() -> PyClassInitializer<Self> {
        PyClassInitializer::from(SubClass::new()).add_subclass(SubSubClass { val3: 20 })
    }

    fn method3(self_: PyRef<'_, Self>) -> PyResult<usize> {
        let v = self_.val3;
        let super_ = self_.into_super(); // Get PyRef<'_, SubClass>
        SubClass::method2(super_).map(|x| x * v)
    }

    #[staticmethod]
    fn factory_method(py: Python<'_>, val: usize) -> PyResult<PyObject> {
        let base = PyClassInitializer::from(BaseClass::new());
        let sub = base.add_subclass(SubClass { val2: val });
        if val % 2 == 0 {
            Ok(Py::new(py, sub)?.to_object(py))
        } else {
            let sub_sub = sub.add_subclass(SubSubClass { val3: val });
            Ok(Py::new(py, sub_sub)?.to_object(py))
        }
    }
}
Python::with_gil(|py| {
    let subsub = pyo3::Py::new(py, SubSubClass::new()).unwrap();
    pyo3::py_run!(py, subsub, "assert subsub.method3() == 3000");
    let subsub = SubSubClass::factory_method(py, 2).unwrap();
    let subsubsub = SubSubClass::factory_method(py, 3).unwrap();
    let cls = py.get_type_bound::<SubSubClass>();
    pyo3::py_run!(py, subsub cls, "assert not isinstance(subsub, cls)");
    pyo3::py_run!(py, subsubsub cls, "assert isinstance(subsubsub, cls)");
});

You can inherit native types such as PyDict, if they implement PySizedLayout. This is not supported when building for the Python limited API (aka the abi3 feature of PyO3).

To convert between the Rust type and its native base class, you can take slf as a Python object. To access the Rust fields use slf.borrow() or slf.borrow_mut(), and to access the base class use slf.downcast::<BaseClass>().

#[cfg(not(Py_LIMITED_API))] {
use pyo3::prelude::*;
use pyo3::types::PyDict;
use std::collections::HashMap;

#[pyclass(extends=PyDict)]
#[derive(Default)]
struct DictWithCounter {
    counter: HashMap<String, usize>,
}

#[pymethods]
impl DictWithCounter {
    #[new]
    fn new() -> Self {
        Self::default()
    }

    fn set(slf: &Bound<'_, Self>, key: String, value: Bound<'_, PyAny>) -> PyResult<()> {
        slf.borrow_mut().counter.entry(key.clone()).or_insert(0);
        let dict = slf.downcast::<PyDict>()?;
        dict.set_item(key, value)
    }
}
Python::with_gil(|py| {
    let cnt = pyo3::Py::new(py, DictWithCounter::new()).unwrap();
    pyo3::py_run!(py, cnt, "cnt.set('abc', 10); assert cnt['abc'] == 10")
});
}

If SubClass does not provide a base class initialization, the compilation fails.

use pyo3::prelude::*;

#[pyclass]
struct BaseClass {
    val1: usize,
}

#[pyclass(extends=BaseClass)]
struct SubClass {
    val2: usize,
}

#[pymethods]
impl SubClass {
    #[new]
    fn new() -> Self {
        SubClass { val2: 15 }
    }
}

The __new__ constructor of a native base class is called implicitly when creating a new instance from Python. Be sure to accept arguments in the #[new] method that you want the base class to get, even if they are not used in that fn:

#[allow(dead_code)]
#[cfg(not(Py_LIMITED_API))] {
use pyo3::prelude::*;
use pyo3::types::PyDict;

#[pyclass(extends=PyDict)]
struct MyDict {
    private: i32,
}

#[pymethods]
impl MyDict {
    #[new]
    #[pyo3(signature = (*args, **kwargs))]
    fn new(args: &Bound<'_, PyAny>, kwargs: Option<&Bound<'_, PyAny>>) -> Self {
        Self { private: 0 }
    }

    // some custom methods that use `private` here...
}
Python::with_gil(|py| {
    let cls = py.get_type_bound::<MyDict>();
    pyo3::py_run!(py, cls, "cls(a=1, b=2)")
});
}

Here, the args and kwargs allow creating instances of the subclass passing initial items, such as MyDict(item_sequence) or MyDict(a=1, b=2).

Object properties

PyO3 supports two ways to add properties to your #[pyclass]:

• For simple struct fields with no side effects, a #[pyo3(get, set)] attribute can be added directly to the field definition in the #[pyclass].
• For properties which require computation you can define #[getter] and #[setter] functions in the #[pymethods] block.

We'll cover each of these in the following sections.

Object properties using #[pyo3(get, set)]

For simple cases where a member variable is just read and written with no side effects, you can declare getters and setters in your #[pyclass] field definition using the pyo3 attribute, like in the example below:

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    #[pyo3(get, set)]
    num: i32,
}

The above would make the num field available for reading and writing as a self.num Python property. To expose the property with a different name to the field, specify this alongside the rest of the options, e.g. #[pyo3(get, set, name = "custom_name")].

Properties can be readonly or writeonly by using just #[pyo3(get)] or #[pyo3(set)] respectively.

To use these annotations, your field type must implement some conversion traits:

• For get the field type must implement both IntoPy<PyObject> and Clone.
• For set the field type must implement FromPyObject.

For example, implementations of those traits are provided for the Cell type, if the inner type also implements the trait. This means you can use #[pyo3(get, set)] on fields wrapped in a Cell.

Object properties using #[getter] and #[setter]

For cases which don't satisfy the #[pyo3(get, set)] trait requirements, or need side effects, descriptor methods can be defined in a #[pymethods] impl block.

This is done using the #[getter] and #[setter] attributes, like in the example below:

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    num: i32,
}

#[pymethods]
impl MyClass {
    #[getter]
    fn num(&self) -> PyResult<i32> {
        Ok(self.num)
    }
}

A getter or setter's function name is used as the property name by default. There are several ways how to override the name.

If a function name starts with get_ or set_ for getter or setter respectively, the descriptor name becomes the function name with this prefix removed. This is also useful in case of Rust keywords like type (raw identifiers can be used since Rust 2018).

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    num: i32,
}
#[pymethods]
impl MyClass {
    #[getter]
    fn get_num(&self) -> PyResult<i32> {
        Ok(self.num)
    }

    #[setter]
    fn set_num(&mut self, value: i32) -> PyResult<()> {
        self.num = value;
        Ok(())
    }
}

In this case, a property num is defined and available from Python code as self.num.

Both the #[getter] and #[setter] attributes accept one parameter. If this parameter is specified, it is used as the property name, i.e.

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
   num: i32,
}
#[pymethods]
impl MyClass {
    #[getter(number)]
    fn num(&self) -> PyResult<i32> {
        Ok(self.num)
    }

    #[setter(number)]
    fn set_num(&mut self, value: i32) -> PyResult<()> {
        self.num = value;
        Ok(())
    }
}

In this case, the property number is defined and available from Python code as self.number.

Attributes defined by #[setter] or #[pyo3(set)] will always raise AttributeError on del operations. Support for defining custom del behavior is tracked in #1778.

Instance methods

To define a Python compatible method, an impl block for your struct has to be annotated with the #[pymethods] attribute. PyO3 generates Python compatible wrappers for all functions in this block with some variations, like descriptors, class method static methods, etc.

Since Rust allows any number of impl blocks, you can easily split methods between those accessible to Python (and Rust) and those accessible only to Rust. However to have multiple #[pymethods]-annotated impl blocks for the same struct you must enable the multiple-pymethods feature of PyO3.

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    num: i32,
}
#[pymethods]
impl MyClass {
    fn method1(&self) -> PyResult<i32> {
        Ok(10)
    }

    fn set_method(&mut self, value: i32) -> PyResult<()> {
        self.num = value;
        Ok(())
    }
}

Calls to these methods are protected by the GIL, so both &self and &mut self can be used. The return type must be PyResult<T> or T for some T that implements IntoPy<PyObject>; the latter is allowed if the method cannot raise Python exceptions.

A Python parameter can be specified as part of method signature, in this case the py argument gets injected by the method wrapper, e.g.

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
#[allow(dead_code)]
    num: i32,
}
#[pymethods]
impl MyClass {
    fn method2(&self, py: Python<'_>) -> PyResult<i32> {
        Ok(10)
    }
}

From the Python perspective, the method2 in this example does not accept any arguments.

Class methods

To create a class method for a custom class, the method needs to be annotated with the #[classmethod] attribute. This is the equivalent of the Python decorator @classmethod.

use pyo3::prelude::*;
use pyo3::types::PyType;
#[pyclass]
struct MyClass {
    #[allow(dead_code)]
    num: i32,
}
#[pymethods]
impl MyClass {
    #[classmethod]
    fn cls_method(cls: &Bound<'_, PyType>) -> PyResult<i32> {
        Ok(10)
    }
}

Declares a class method callable from Python.

• The first parameter is the type object of the class on which the method is called. This may be the type object of a derived class.
• The first parameter implicitly has type &Bound<'_, PyType>.
• For details on parameter-list, see the documentation of Method arguments section.
• The return type must be PyResult<T> or T for some T that implements IntoPy<PyObject>.

Constructors which accept a class argument

To create a constructor which takes a positional class argument, you can combine the #[classmethod] and #[new] modifiers:

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyType;
#[pyclass]
struct BaseClass(PyObject);

#[pymethods]
impl BaseClass {
    #[new]
    #[classmethod]
    fn py_new(cls: &Bound<'_, PyType>) -> PyResult<Self> {
        // Get an abstract attribute (presumably) declared on a subclass of this class.
        let subclass_attr: Bound<'_, PyAny> = cls.getattr("a_class_attr")?;
        Ok(Self(subclass_attr.unbind()))
    }
}

Static methods

To create a static method for a custom class, the method needs to be annotated with the #[staticmethod] attribute. The return type must be T or PyResult<T> for some T that implements IntoPy<PyObject>.

use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    #[allow(dead_code)]
    num: i32,
}
#[pymethods]
impl MyClass {
    #[staticmethod]
    fn static_method(param1: i32, param2: &str) -> PyResult<i32> {
        Ok(10)
    }
}

Class attributes

To create a class attribute (also called class variable), a method without any arguments can be annotated with the #[classattr] attribute.

use pyo3::prelude::*;
#[pyclass]
struct MyClass {}
#[pymethods]
impl MyClass {
    #[classattr]
    fn my_attribute() -> String {
        "hello".to_string()
    }
}

Python::with_gil(|py| {
    let my_class = py.get_type_bound::<MyClass>();
    pyo3::py_run!(py, my_class, "assert my_class.my_attribute == 'hello'")
});

Note: if the method has a Result return type and returns an Err, PyO3 will panic during class creation.

If the class attribute is defined with const code only, one can also annotate associated constants:

use pyo3::prelude::*;
#[pyclass]
struct MyClass {}
#[pymethods]
impl MyClass {
    #[classattr]
    const MY_CONST_ATTRIBUTE: &'static str = "foobar";
}

Classes as function arguments

Free functions defined using #[pyfunction] interact with classes through the same mechanisms as the self parameters of instance methods, i.e. they can take GIL-bound references, GIL-bound reference wrappers or GIL-indepedent references:

#![allow(dead_code)]
use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    my_field: i32,
}

// Take a reference when the underlying `Bound` is irrelevant.
#[pyfunction]
fn increment_field(my_class: &mut MyClass) {
    my_class.my_field += 1;
}

// Take a reference wrapper when borrowing should be automatic,
// but interaction with the underlying `Bound` is desired.
#[pyfunction]
fn print_field(my_class: PyRef<'_, MyClass>) {
    println!("{}", my_class.my_field);
}

// Take a reference to the underlying Bound
// when borrowing needs to be managed manually.
#[pyfunction]
fn increment_then_print_field(my_class: &Bound<'_, MyClass>) {
    my_class.borrow_mut().my_field += 1;

    println!("{}", my_class.borrow().my_field);
}

// Take a GIL-indepedent reference when you want to store the reference elsewhere.
#[pyfunction]
fn print_refcnt(my_class: Py<MyClass>, py: Python<'_>) {
    println!("{}", my_class.get_refcnt(py));
}

Classes can also be passed by value if they can be cloned, i.e. they automatically implement FromPyObject if they implement Clone, e.g. via #[derive(Clone)]:

#![allow(dead_code)]
use pyo3::prelude::*;
#[pyclass]
#[derive(Clone)]
struct MyClass {
    my_field: Box<i32>,
}

#[pyfunction]
fn dissamble_clone(my_class: MyClass) {
    let MyClass { mut my_field } = my_class;
    *my_field += 1;
}

Note that #[derive(FromPyObject)] on a class is usually not useful as it tries to construct a new Rust value by filling in the fields by looking up attributes of any given Python value.

Method arguments

Similar to #[pyfunction], the #[pyo3(signature = (...))] attribute can be used to specify the way that #[pymethods] accept arguments. Consult the documentation for function signatures to see the parameters this attribute accepts.

The following example defines a class MyClass with a method method. This method has a signature that sets default values for num and name, and indicates that py_args should collect all extra positional arguments and py_kwargs all extra keyword arguments:

use pyo3::prelude::*;
use pyo3::types::{PyDict, PyTuple};

#[pyclass]
struct MyClass {
    num: i32,
}
#[pymethods]
impl MyClass {
    #[new]
    #[pyo3(signature = (num=-1))]
    fn new(num: i32) -> Self {
        MyClass { num }
    }

    #[pyo3(signature = (num=10, *py_args, name="Hello", **py_kwargs))]
    fn method(
        &mut self,
        num: i32,
        py_args: &Bound<'_, PyTuple>,
        name: &str,
        py_kwargs: Option<&Bound<'_, PyDict>>,
    ) -> String {
        let num_before = self.num;
        self.num = num;
        format!(
            "num={} (was previously={}), py_args={:?}, name={}, py_kwargs={:?} ",
            num, num_before, py_args, name, py_kwargs,
        )
    }
}

In Python, this might be used like:

>>> import mymodule
>>> mc = mymodule.MyClass()
>>> print(mc.method(44, False, "World", 666, x=44, y=55))
py_args=('World', 666), py_kwargs=Some({'x': 44, 'y': 55}), name=Hello, num=44, num_before=-1
>>> print(mc.method(num=-1, name="World"))
py_args=(), py_kwargs=None, name=World, num=-1, num_before=44

The #[pyo3(text_signature = "...") option for #[pyfunction] also works for #[pymethods].

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyType;

#[pyclass]
struct MyClass {}

#[pymethods]
impl MyClass {
    #[new]
    #[pyo3(text_signature = "(c, d)")]
    fn new(c: i32, d: &str) -> Self {
        Self {}
    }
    // the self argument should be written $self
    #[pyo3(text_signature = "($self, e, f)")]
    fn my_method(&self, e: i32, f: i32) -> i32 {
        e + f
    }
    // similarly for classmethod arguments, use $cls
    #[classmethod]
    #[pyo3(text_signature = "($cls, e, f)")]
    fn my_class_method(cls: &Bound<'_, PyType>, e: i32, f: i32) -> i32 {
        e + f
    }
    #[staticmethod]
    #[pyo3(text_signature = "(e, f)")]
    fn my_static_method(e: i32, f: i32) -> i32 {
        e + f
    }
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let inspect = PyModule::import_bound(py, "inspect")?.getattr("signature")?;
        let module = PyModule::new_bound(py, "my_module")?;
        module.add_class::<MyClass>()?;
        let class = module.getattr("MyClass")?;

        if cfg!(not(Py_LIMITED_API)) || py.version_info() >= (3, 10)  {
            let doc: String = class.getattr("__doc__")?.extract()?;
            assert_eq!(doc, "");

            let sig: String = inspect
                .call1((&class,))?
                .call_method0("__str__")?
                .extract()?;
            assert_eq!(sig, "(c, d)");
        } else {
            let doc: String = class.getattr("__doc__")?.extract()?;
            assert_eq!(doc, "");

            inspect.call1((&class,)).expect_err("`text_signature` on classes is not compatible with compilation in `abi3` mode until Python 3.10 or greater");
         }

        {
            let method = class.getattr("my_method")?;

            assert!(method.getattr("__doc__")?.is_none());

            let sig: String = inspect
                .call1((method,))?
                .call_method0("__str__")?
                .extract()?;
            assert_eq!(sig, "(self, /, e, f)");
        }

        {
            let method = class.getattr("my_class_method")?;

            assert!(method.getattr("__doc__")?.is_none());

            let sig: String = inspect
                .call1((method,))?
                .call_method0("__str__")?
                .extract()?;
            assert_eq!(sig, "(e, f)");  // inspect.signature skips the $cls arg
        }

        {
            let method = class.getattr("my_static_method")?;

            assert!(method.getattr("__doc__")?.is_none());

            let sig: String = inspect
                .call1((method,))?
                .call_method0("__str__")?
                .extract()?;
            assert_eq!(sig, "(e, f)");
        }

        Ok(())
    })
}

Note that text_signature on #[new] is not compatible with compilation in abi3 mode until Python 3.10 or greater.

Method receivers and lifetime elision

PyO3 supports writing instance methods using the normal method receivers for shared &self and unique &mut self references. This interacts with lifetime elision insofar as the lifetime of a such a receiver is assigned to all elided output lifetime parameters.

This is a good default for general Rust code where return values are more likely to borrow from the receiver than from the other arguments, if they contain any lifetimes at all. However, when returning bound references Bound<'py, T> in PyO3-based code, the GIL lifetime 'py should usually be derived from a GIL token py: Python<'py> passed as an argument instead of the receiver.

Specifically, signatures like

fn frobnicate(&self, py: Python) -> Bound<Foo>;

will not work as they are inferred as

fn frobnicate<'a, 'py>(&'a self, py: Python<'py>) -> Bound<'a, Foo>;

instead of the intended

fn frobnicate<'a, 'py>(&'a self, py: Python<'py>) -> Bound<'py, Foo>;

and should usually be written as

fn frobnicate<'py>(&self, py: Python<'py>) -> Bound<'py, Foo>;

The same problem does not exist for #[pyfunction]s as the special case for receiver lifetimes does not apply and indeed a signature like

fn frobnicate(bar: &Bar, py: Python) -> Bound<Foo>;

will yield compiler error E0106 "missing lifetime specifier".

#[pyclass] enums

Enum support in PyO3 comes in two flavors, depending on what kind of variants the enum has: simple and complex.

Simple enums

A simple enum (a.k.a. C-like enum) has only unit variants.

PyO3 adds a class attribute for each variant, so you can access them in Python without defining #[new]. PyO3 also provides default implementations of __richcmp__ and __int__, so they can be compared using ==:

use pyo3::prelude::*;
#[pyclass]
enum MyEnum {
    Variant,
    OtherVariant,
}

Python::with_gil(|py| {
    let x = Py::new(py, MyEnum::Variant).unwrap();
    let y = Py::new(py, MyEnum::OtherVariant).unwrap();
    let cls = py.get_type_bound::<MyEnum>();
    pyo3::py_run!(py, x y cls, r#"
        assert x == cls.Variant
        assert y == cls.OtherVariant
        assert x != y
    "#)
})

You can also convert your simple enums into int:

use pyo3::prelude::*;
#[pyclass]
enum MyEnum {
    Variant,
    OtherVariant = 10,
}

Python::with_gil(|py| {
    let cls = py.get_type_bound::<MyEnum>();
    let x = MyEnum::Variant as i32; // The exact value is assigned by the compiler.
    pyo3::py_run!(py, cls x, r#"
        assert int(cls.Variant) == x
        assert int(cls.OtherVariant) == 10
        assert cls.OtherVariant == 10  # You can also compare against int.
        assert 10 == cls.OtherVariant
    "#)
})

PyO3 also provides __repr__ for enums:

use pyo3::prelude::*;
#[pyclass]
enum MyEnum{
    Variant,
    OtherVariant,
}

Python::with_gil(|py| {
    let cls = py.get_type_bound::<MyEnum>();
    let x = Py::new(py, MyEnum::Variant).unwrap();
    pyo3::py_run!(py, cls x, r#"
        assert repr(x) == 'MyEnum.Variant'
        assert repr(cls.OtherVariant) == 'MyEnum.OtherVariant'
    "#)
})

All methods defined by PyO3 can be overridden. For example here's how you override __repr__:

use pyo3::prelude::*;
#[pyclass]
enum MyEnum {
    Answer = 42,
}

#[pymethods]
impl MyEnum {
    fn __repr__(&self) -> &'static str {
        "42"
    }
}

Python::with_gil(|py| {
    let cls = py.get_type_bound::<MyEnum>();
    pyo3::py_run!(py, cls, "assert repr(cls.Answer) == '42'")
})

Enums and their variants can also be renamed using #[pyo3(name)].

use pyo3::prelude::*;
#[pyclass(name = "RenamedEnum")]
enum MyEnum {
    #[pyo3(name = "UPPERCASE")]
    Variant,
}

Python::with_gil(|py| {
    let x = Py::new(py, MyEnum::Variant).unwrap();
    let cls = py.get_type_bound::<MyEnum>();
    pyo3::py_run!(py, x cls, r#"
        assert repr(x) == 'RenamedEnum.UPPERCASE'
        assert x == cls.UPPERCASE
    "#)
})

You may not use enums as a base class or let enums inherit from other classes.

use pyo3::prelude::*;
#[pyclass(subclass)]
enum BadBase {
    Var1,
}

use pyo3::prelude::*;

#[pyclass(subclass)]
struct Base;

#[pyclass(extends=Base)]
enum BadSubclass {
    Var1,
}

#[pyclass] enums are currently not interoperable with IntEnum in Python.

Complex enums

An enum is complex if it has any non-unit (struct or tuple) variants.

Currently PyO3 supports only struct variants in a complex enum. Support for unit and tuple variants is planned.

PyO3 adds a class attribute for each variant, which may be used to construct values and in match patterns. PyO3 also provides getter methods for all fields of each variant.

use pyo3::prelude::*;
#[pyclass]
enum Shape {
    Circle { radius: f64 },
    Rectangle { width: f64, height: f64 },
    RegularPolygon { side_count: u32, radius: f64 },
    Nothing { },
}

#[cfg(Py_3_10)]
Python::with_gil(|py| {
    let circle = Shape::Circle { radius: 10.0 }.into_py(py);
    let square = Shape::RegularPolygon { side_count: 4, radius: 10.0 }.into_py(py);
    let cls = py.get_type_bound::<Shape>();
    pyo3::py_run!(py, circle square cls, r#"
        assert isinstance(circle, cls)
        assert isinstance(circle, cls.Circle)
        assert circle.radius == 10.0

        assert isinstance(square, cls)
        assert isinstance(square, cls.RegularPolygon)
        assert square.side_count == 4
        assert square.radius == 10.0

        def count_vertices(cls, shape):
            match shape:
                case cls.Circle():
                    return 0
                case cls.Rectangle():
                    return 4
                case cls.RegularPolygon(side_count=n):
                    return n
                case cls.Nothing():
                    return 0

        assert count_vertices(cls, circle) == 0
        assert count_vertices(cls, square) == 4
    "#)
})

WARNING: Py::new and .into_py are currently inconsistent. Note how the constructed value is not an instance of the specific variant. For this reason, constructing values is only recommended using .into_py.

use pyo3::prelude::*;
#[pyclass]
enum MyEnum {
    Variant { i: i32 },
}

Python::with_gil(|py| {
    let x = Py::new(py, MyEnum::Variant { i: 42 }).unwrap();
    let cls = py.get_type_bound::<MyEnum>();
    pyo3::py_run!(py, x cls, r#"
        assert isinstance(x, cls)
        assert not isinstance(x, cls.Variant)
    "#)
})

Implementation details

The #[pyclass] macros rely on a lot of conditional code generation: each #[pyclass] can optionally have a #[pymethods] block.

To support this flexibility the #[pyclass] macro expands to a blob of boilerplate code which sets up the structure for "dtolnay specialization". This implementation pattern enables the Rust compiler to use #[pymethods] implementations when they are present, and fall back to default (empty) definitions when they are not.

This simple technique works for the case when there is zero or one implementations. To support multiple #[pymethods] for a #[pyclass] (in the multiple-pymethods feature), a registry mechanism provided by the inventory crate is used instead. This collects impls at library load time, but isn't supported on all platforms. See inventory: how it works for more details.

The #[pyclass] macro expands to roughly the code seen below. The PyClassImplCollector is the type used internally by PyO3 for dtolnay specialization:

#[cfg(not(feature = "multiple-pymethods"))] {
use pyo3::prelude::*;
// Note: the implementation differs slightly with the `multiple-pymethods` feature enabled.
struct MyClass {
    #[allow(dead_code)]
    num: i32,
}

impl pyo3::types::DerefToPyAny for MyClass {}

#[allow(deprecated)]
unsafe impl pyo3::type_object::HasPyGilRef for MyClass {
    type AsRefTarget = pyo3::PyCell<Self>;
}
unsafe impl pyo3::type_object::PyTypeInfo for MyClass {
    const NAME: &'static str = "MyClass";
    const MODULE: ::std::option::Option<&'static str> = ::std::option::Option::None;
    #[inline]
    fn type_object_raw(py: pyo3::Python<'_>) -> *mut pyo3::ffi::PyTypeObject {
        <Self as pyo3::impl_::pyclass::PyClassImpl>::lazy_type_object()
            .get_or_init(py)
            .as_type_ptr()
    }
}

impl pyo3::PyClass for MyClass {
    type Frozen = pyo3::pyclass::boolean_struct::False;
}

impl<'a, 'py> pyo3::impl_::extract_argument::PyFunctionArgument<'a, 'py> for &'a MyClass
{
    type Holder = ::std::option::Option<pyo3::PyRef<'py, MyClass>>;

    #[inline]
    fn extract(obj: &'a pyo3::Bound<'py, PyAny>, holder: &'a mut Self::Holder) -> pyo3::PyResult<Self> {
        pyo3::impl_::extract_argument::extract_pyclass_ref(obj, holder)
    }
}

impl<'a, 'py> pyo3::impl_::extract_argument::PyFunctionArgument<'a, 'py> for &'a mut MyClass
{
    type Holder = ::std::option::Option<pyo3::PyRefMut<'py, MyClass>>;

    #[inline]
    fn extract(obj: &'a pyo3::Bound<'py, PyAny>, holder: &'a mut Self::Holder) -> pyo3::PyResult<Self> {
        pyo3::impl_::extract_argument::extract_pyclass_ref_mut(obj, holder)
    }
}

impl pyo3::IntoPy<PyObject> for MyClass {
    fn into_py(self, py: pyo3::Python<'_>) -> pyo3::PyObject {
        pyo3::IntoPy::into_py(pyo3::Py::new(py, self).unwrap(), py)
    }
}

impl pyo3::impl_::pyclass::PyClassImpl for MyClass {
    const IS_BASETYPE: bool = false;
    const IS_SUBCLASS: bool = false;
    const IS_MAPPING: bool = false;
    const IS_SEQUENCE: bool = false;
    type BaseType = PyAny;
    type ThreadChecker = pyo3::impl_::pyclass::SendablePyClass<MyClass>;
    type PyClassMutability = <<pyo3::PyAny as pyo3::impl_::pyclass::PyClassBaseType>::PyClassMutability as pyo3::impl_::pycell::PyClassMutability>::MutableChild;
    type Dict = pyo3::impl_::pyclass::PyClassDummySlot;
    type WeakRef = pyo3::impl_::pyclass::PyClassDummySlot;
    type BaseNativeType = pyo3::PyAny;

    fn items_iter() -> pyo3::impl_::pyclass::PyClassItemsIter {
        use pyo3::impl_::pyclass::*;
        let collector = PyClassImplCollector::<MyClass>::new();
        static INTRINSIC_ITEMS: PyClassItems = PyClassItems { slots: &[], methods: &[] };
        PyClassItemsIter::new(&INTRINSIC_ITEMS, collector.py_methods())
    }

    fn lazy_type_object() -> &'static pyo3::impl_::pyclass::LazyTypeObject<MyClass> {
        use pyo3::impl_::pyclass::LazyTypeObject;
        static TYPE_OBJECT: LazyTypeObject<MyClass> = LazyTypeObject::new();
        &TYPE_OBJECT
    }

    fn doc(py: Python<'_>) -> pyo3::PyResult<&'static ::std::ffi::CStr> {
        use pyo3::impl_::pyclass::*;
        static DOC: pyo3::sync::GILOnceCell<::std::borrow::Cow<'static, ::std::ffi::CStr>> = pyo3::sync::GILOnceCell::new();
        DOC.get_or_try_init(py, || {
            let collector = PyClassImplCollector::<Self>::new();
            build_pyclass_doc(<MyClass as pyo3::PyTypeInfo>::NAME, "\0", collector.new_text_signature())
        }).map(::std::ops::Deref::deref)
    }
}

Python::with_gil(|py| {
    let cls = py.get_type_bound::<MyClass>();
    pyo3::py_run!(py, cls, "assert cls.__name__ == 'MyClass'")
});
}

Magic methods and slots

Python's object model defines several protocols for different object behavior, such as the sequence, mapping, and number protocols. You may be familiar with implementing these protocols in Python classes by "magic" methods, such as __str__ or __repr__. Because of the double-underscores surrounding their name, these are also known as "dunder" methods.

In the Python C-API which PyO3 is implemented upon, many of these magic methods have to be placed into special "slots" on the class type object, as covered in the previous section.

If a function name in #[pymethods] is a recognised magic method, it will be automatically placed into the correct slot in the Python type object. The function name is taken from the usual rules for naming #[pymethods]: the #[pyo3(name = "...")] attribute is used if present, otherwise the Rust function name is used.

The magic methods handled by PyO3 are very similar to the standard Python ones on this page - in particular they are the subset which have slots as defined here. Some of the slots do not have a magic method in Python, which leads to a few additional magic methods defined only in PyO3:

• Magic methods for garbage collection
• Magic methods for the buffer protocol

When PyO3 handles a magic method, a couple of changes apply compared to other #[pymethods]:

• The Rust function signature is restricted to match the magic method.
• The #[pyo3(signature = (...)] and #[pyo3(text_signature = "...")] attributes are not allowed.

The following sections list of all magic methods PyO3 currently handles. The given signatures should be interpreted as follows:

• All methods take a receiver as first argument, shown as <self>. It can be &self, &mut self or a Bound reference like self_: PyRef<'_, Self> and self_: PyRefMut<'_, Self>, as described here.
• An optional Python<'py> argument is always allowed as the first argument.
• Return values can be optionally wrapped in PyResult.
• object means that any type is allowed that can be extracted from a Python object (if argument) or converted to a Python object (if return value).
• Other types must match what's given, e.g. pyo3::basic::CompareOp for __richcmp__'s second argument.
• For the comparison and arithmetic methods, extraction errors are not propagated as exceptions, but lead to a return of NotImplemented.
• For some magic methods, the return values are not restricted by PyO3, but checked by the Python interpreter. For example, __str__ needs to return a string object. This is indicated by object (Python type).

Basic object customization

• __str__(<self>) -> object (str)

• __repr__(<self>) -> object (str)

• __hash__(<self>) -> isize

  Objects that compare equal must have the same hash value. Any type up to 64 bits may be returned instead of isize, PyO3 will convert to an isize automatically (wrapping unsigned types like u64 and usize).

  Disabling Python's default hash

  By default, all `#[pyclass]` types have a default hash implementation from Python. Types which should not be hashable can override this by setting `__hash__` to `None`. This is the same mechanism as for a pure-Python class. This is done like so:

  use pyo3::prelude::*;
  
  #[pyclass]
  struct NotHashable {}
  
  #[pymethods]
  impl NotHashable {
      #[classattr]
      const __hash__: Option<PyObject> = None;
  }

• __lt__(<self>, object) -> object

• __le__(<self>, object) -> object

• __eq__(<self>, object) -> object

• __ne__(<self>, object) -> object

• __gt__(<self>, object) -> object

• __ge__(<self>, object) -> object

  The implementations of Python's "rich comparison" operators <, <=, ==, !=, > and >= respectively.

  Note that implementing any of these methods will cause Python not to generate a default __hash__ implementation, so consider also implementing __hash__.

  Return type

  The return type will normally be `bool` or `PyResult`, however any Python object can be returned.

• __richcmp__(<self>, object, pyo3::basic::CompareOp) -> object

  Implements Python comparison operations (==, !=, <, <=, >, and >=) in a single method. The CompareOp argument indicates the comparison operation being performed. You can use CompareOp::matches to adapt a Rust std::cmp::Ordering result to the requested comparison.

  This method cannot be implemented in combination with any of __lt__, __le__, __eq__, __ne__, __gt__, or __ge__.

  Note that implementing __richcmp__ will cause Python not to generate a default __hash__ implementation, so consider implementing __hash__ when implementing __richcmp__.

  Return type

  The return type will normally be `PyResult`, but any Python object can be returned.

  If you want to leave some operations unimplemented, you can return py.NotImplemented() for some of the operations:

  use pyo3::class::basic::CompareOp;
  
  use pyo3::prelude::*;
  
  #[pyclass]
  struct Number(i32);
  
  #[pymethods]
  impl Number {
      fn __richcmp__(&self, other: &Self, op: CompareOp, py: Python<'_>) -> PyObject {
          match op {
              CompareOp::Eq => (self.0 == other.0).into_py(py),
              CompareOp::Ne => (self.0 != other.0).into_py(py),
              _ => py.NotImplemented(),
          }
      }
  }

  If the second argument object is not of the type specified in the signature, the generated code will automatically return NotImplemented.

• __getattr__(<self>, object) -> object

• __getattribute__(<self>, object) -> object

  Differences between `__getattr__` and `__getattribute__`

  As in Python, `__getattr__` is only called if the attribute is not found by normal attribute lookup. `__getattribute__`, on the other hand, is called for *every* attribute access. If it wants to access existing attributes on `self`, it needs to be very careful not to introduce infinite recursion, and use `baseclass.__getattribute__()`.

• __setattr__(<self>, value: object) -> ()

• __delattr__(<self>, object) -> ()

  Overrides attribute access.

• __bool__(<self>) -> bool

  Determines the "truthyness" of an object.

• __call__(<self>, ...) -> object - here, any argument list can be defined as for normal pymethods

Iterable objects

Iterators can be defined using these methods:

• __iter__(<self>) -> object
• __next__(<self>) -> Option<object> or IterNextOutput (see details)

Returning None from __next__ indicates that that there are no further items.

Example:

use pyo3::prelude::*;

#[pyclass]
struct MyIterator {
    iter: Box<dyn Iterator<Item = PyObject> + Send>,
}

#[pymethods]
impl MyIterator {
    fn __iter__(slf: PyRef<'_, Self>) -> PyRef<'_, Self> {
        slf
    }
    fn __next__(mut slf: PyRefMut<'_, Self>) -> Option<PyObject> {
        slf.iter.next()
    }
}

In many cases you'll have a distinction between the type being iterated over (i.e. the iterable) and the iterator it provides. In this case, the iterable only needs to implement __iter__() while the iterator must implement both __iter__() and __next__(). For example:

use pyo3::prelude::*;

#[pyclass]
struct Iter {
    inner: std::vec::IntoIter<usize>,
}

#[pymethods]
impl Iter {
    fn __iter__(slf: PyRef<'_, Self>) -> PyRef<'_, Self> {
        slf
    }

    fn __next__(mut slf: PyRefMut<'_, Self>) -> Option<usize> {
        slf.inner.next()
    }
}

#[pyclass]
struct Container {
    iter: Vec<usize>,
}

#[pymethods]
impl Container {
    fn __iter__(slf: PyRef<'_, Self>) -> PyResult<Py<Iter>> {
        let iter = Iter {
            inner: slf.iter.clone().into_iter(),
        };
        Py::new(slf.py(), iter)
    }
}

Python::with_gil(|py| {
    let container = Container { iter: vec![1, 2, 3, 4] };
    let inst = pyo3::Py::new(py, container).unwrap();
    pyo3::py_run!(py, inst, "assert list(inst) == [1, 2, 3, 4]");
    pyo3::py_run!(py, inst, "assert list(iter(iter(inst))) == [1, 2, 3, 4]");
});

For more details on Python's iteration protocols, check out the "Iterator Types" section of the library documentation.

Returning a value from iteration

This guide has so far shown how to use Option<T> to implement yielding values during iteration. In Python a generator can also return a value. To express this in Rust, PyO3 provides the IterNextOutput enum to both Yield values and Return a final value - see its docs for further details and an example.

Awaitable objects

• __await__(<self>) -> object
• __aiter__(<self>) -> object
• __anext__(<self>) -> Option<object> or IterANextOutput

Mapping & Sequence types

The magic methods in this section can be used to implement Python container types. They are two main categories of container in Python: "mappings" such as dict, with arbitrary keys, and "sequences" such as list and tuple, with integer keys.

The Python C-API which PyO3 is built upon has separate "slots" for sequences and mappings. When writing a class in pure Python, there is no such distinction in the implementation - a __getitem__ implementation will fill the slots for both the mapping and sequence forms, for example.

By default PyO3 reproduces the Python behaviour of filling both mapping and sequence slots. This makes sense for the "simple" case which matches Python, and also for sequences, where the mapping slot is used anyway to implement slice indexing.

Mapping types usually will not want the sequence slots filled. Having them filled will lead to outcomes which may be unwanted, such as:

• The mapping type will successfully cast to PySequence. This may lead to consumers of the type handling it incorrectly.
• Python provides a default implementation of __iter__ for sequences, which calls __getitem__ with consecutive positive integers starting from 0 until an IndexError is returned. Unless the mapping only contains consecutive positive integer keys, this __iter__ implementation will likely not be the intended behavior.

Use the #[pyclass(mapping)] annotation to instruct PyO3 to only fill the mapping slots, leaving the sequence ones empty. This will apply to __getitem__, __setitem__, and __delitem__.

Use the #[pyclass(sequence)] annotation to instruct PyO3 to fill the sq_length slot instead of the mp_length slot for __len__. This will help libraries such as numpy recognise the class as a sequence, however will also cause CPython to automatically add the sequence length to any negative indices before passing them to __getitem__. (__getitem__, __setitem__ and __delitem__ mapping slots are still used for sequences, for slice operations.)

• __len__(<self>) -> usize

  Implements the built-in function len().

• __contains__(<self>, object) -> bool

  Implements membership test operators. Should return true if item is in self, false otherwise. For objects that don’t define __contains__(), the membership test simply traverses the sequence until it finds a match.

  Disabling Python's default contains

  By default, all #[pyclass] types with an __iter__ method support a default implementation of the in operator. Types which do not want this can override this by setting __contains__ to None. This is the same mechanism as for a pure-Python class. This is done like so:

  use pyo3::prelude::*;
  
  #[pyclass]
  struct NoContains {}
  
  #[pymethods]
  impl NoContains {
      #[classattr]
      const __contains__: Option<PyObject> = None;
  }

• __getitem__(<self>, object) -> object

  Implements retrieval of the self[a] element.

  Note: Negative integer indexes are not handled specially by PyO3. However, for classes with #[pyclass(sequence)], when a negative index is accessed via PySequence::get_item, the underlying C API already adjusts the index to be positive.

• __setitem__(<self>, object, object) -> ()

  Implements assignment to the self[a] element. Should only be implemented if elements can be replaced.

  Same behavior regarding negative indices as for __getitem__.

• __delitem__(<self>, object) -> ()

  Implements deletion of the self[a] element. Should only be implemented if elements can be deleted.

  Same behavior regarding negative indices as for __getitem__.

• fn __concat__(&self, other: impl FromPyObject) -> PyResult<impl ToPyObject>

  Concatenates two sequences. Used by the + operator, after trying the numeric addition via the __add__ and __radd__ methods.

• fn __repeat__(&self, count: isize) -> PyResult<impl ToPyObject>

  Repeats the sequence count times. Used by the * operator, after trying the numeric multiplication via the __mul__ and __rmul__ methods.

• fn __inplace_concat__(&self, other: impl FromPyObject) -> PyResult<impl ToPyObject>

  Concatenates two sequences. Used by the += operator, after trying the numeric addition via the __iadd__ method.

• fn __inplace_repeat__(&self, count: isize) -> PyResult<impl ToPyObject>

  Concatenates two sequences. Used by the *= operator, after trying the numeric multiplication via the __imul__ method.

Descriptors

• __get__(<self>, object, object) -> object
• __set__(<self>, object, object) -> ()
• __delete__(<self>, object) -> ()

Numeric types

Binary arithmetic operations (+, -, *, @, /, //, %, divmod(), pow() and **, <<, >>, &, ^, and |) and their reflected versions:

(If the object is not of the type specified in the signature, the generated code will automatically return NotImplemented.)

• __add__(<self>, object) -> object
• __radd__(<self>, object) -> object
• __sub__(<self>, object) -> object
• __rsub__(<self>, object) -> object
• __mul__(<self>, object) -> object
• __rmul__(<self>, object) -> object
• __matmul__(<self>, object) -> object
• __rmatmul__(<self>, object) -> object
• __floordiv__(<self>, object) -> object
• __rfloordiv__(<self>, object) -> object
• __truediv__(<self>, object) -> object
• __rtruediv__(<self>, object) -> object
• __divmod__(<self>, object) -> object
• __rdivmod__(<self>, object) -> object
• __mod__(<self>, object) -> object
• __rmod__(<self>, object) -> object
• __lshift__(<self>, object) -> object
• __rlshift__(<self>, object) -> object
• __rshift__(<self>, object) -> object
• __rrshift__(<self>, object) -> object
• __and__(<self>, object) -> object
• __rand__(<self>, object) -> object
• __xor__(<self>, object) -> object
• __rxor__(<self>, object) -> object
• __or__(<self>, object) -> object
• __ror__(<self>, object) -> object
• __pow__(<self>, object, object) -> object
• __rpow__(<self>, object, object) -> object

In-place assignment operations (+=, -=, *=, @=, /=, //=, %=, **=, <<=, >>=, &=, ^=, |=):

• __iadd__(<self>, object) -> ()
• __isub__(<self>, object) -> ()
• __imul__(<self>, object) -> ()
• __imatmul__(<self>, object) -> ()
• __itruediv__(<self>, object) -> ()
• __ifloordiv__(<self>, object) -> ()
• __imod__(<self>, object) -> ()
• __ipow__(<self>, object, object) -> ()
• __ilshift__(<self>, object) -> ()
• __irshift__(<self>, object) -> ()
• __iand__(<self>, object) -> ()
• __ixor__(<self>, object) -> ()
• __ior__(<self>, object) -> ()

Unary operations (-, +, abs() and ~):

• __pos__(<self>) -> object
• __neg__(<self>) -> object
• __abs__(<self>) -> object
• __invert__(<self>) -> object

Coercions:

• __index__(<self>) -> object (int)
• __int__(<self>) -> object (int)
• __float__(<self>) -> object (float)

Buffer objects

• __getbuffer__(<self>, *mut ffi::Py_buffer, flags) -> ()
• __releasebuffer__(<self>, *mut ffi::Py_buffer) -> () Errors returned from __releasebuffer__ will be sent to sys.unraiseablehook. It is strongly advised to never return an error from __releasebuffer__, and if it really is necessary, to make best effort to perform any required freeing operations before returning. __releasebuffer__ will not be called a second time; anything not freed will be leaked.

Garbage Collector Integration

If your type owns references to other Python objects, you will need to integrate with Python's garbage collector so that the GC is aware of those references. To do this, implement the two methods __traverse__ and __clear__. These correspond to the slots tp_traverse and tp_clear in the Python C API. __traverse__ must call visit.call() for each reference to another Python object. __clear__ must clear out any mutable references to other Python objects (thus breaking reference cycles). Immutable references do not have to be cleared, as every cycle must contain at least one mutable reference.

• __traverse__(<self>, pyo3::class::gc::PyVisit<'_>) -> Result<(), pyo3::class::gc::PyTraverseError>
• __clear__(<self>) -> ()

Example:

use pyo3::prelude::*;
use pyo3::PyTraverseError;
use pyo3::gc::PyVisit;

#[pyclass]
struct ClassWithGCSupport {
    obj: Option<PyObject>,
}

#[pymethods]
impl ClassWithGCSupport {
    fn __traverse__(&self, visit: PyVisit<'_>) -> Result<(), PyTraverseError> {
        if let Some(obj) = &self.obj {
            visit.call(obj)?
        }
        Ok(())
    }

    fn __clear__(&mut self) {
        // Clear reference, this decrements ref counter.
        self.obj = None;
    }
}

Usually, an implementation of __traverse__ should do nothing but calls to visit.call. Most importantly, safe access to the GIL is prohibited inside implementations of __traverse__, i.e. Python::with_gil will panic.

Note: these methods are part of the C API, PyPy does not necessarily honor them. If you are building for PyPy you should measure memory consumption to make sure you do not have runaway memory growth. See this issue on the PyPy bug tracker.

Basic object customization

Recall the Number class from the previous chapter:

#![allow(dead_code)]
use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    #[new]
    fn new(value: i32) -> Self {
        Self(value)
    }
}

#[pymodule]
fn my_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_class::<Number>()?;
    Ok(())
}

At this point Python code can import the module, access the class and create class instances - but nothing else.

from my_module import Number

n = Number(5)
print(n)

<builtins.Number object at 0x000002B4D185D7D0>

String representations

It can't even print an user-readable representation of itself! We can fix that by defining the __repr__ and __str__ methods inside a #[pymethods] block. We do this by accessing the value contained inside Number.

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    // For `__repr__` we want to return a string that Python code could use to recreate
    // the `Number`, like `Number(5)` for example.
    fn __repr__(&self) -> String {
        // We use the `format!` macro to create a string. Its first argument is a
        // format string, followed by any number of parameters which replace the
        // `{}`'s in the format string.
        //
        //                       👇 Tuple field access in Rust uses a dot
        format!("Number({})", self.0)
    }
    // `__str__` is generally used to create an "informal" representation, so we
    // just forward to `i32`'s `ToString` trait implementation to print a bare number.
    fn __str__(&self) -> String {
        self.0.to_string()
    }
}

Accessing the class name

In the __repr__, we used a hard-coded class name. This is sometimes not ideal, because if the class is subclassed in Python, we would like the repr to reflect the subclass name. This is typically done in Python code by accessing self.__class__.__name__. In order to be able to access the Python type information and the Rust struct, we need to use a Bound as the self argument.

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __repr__(slf: &Bound<'_, Self>) -> PyResult<String> {
        // This is the equivalent of `self.__class__.__name__` in Python.
        let class_name: String = slf.get_type().qualname()?;
        // To access fields of the Rust struct, we need to borrow the `PyCell`.
        Ok(format!("{}({})", class_name, slf.borrow().0))
    }
}

Hashing

Let's also implement hashing. We'll just hash the i32. For that we need a Hasher. The one provided by std is DefaultHasher, which uses the SipHash algorithm.

use std::collections::hash_map::DefaultHasher;

// Required to call the `.hash` and `.finish` methods, which are defined on traits.
use std::hash::{Hash, Hasher};

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __hash__(&self) -> u64 {
        let mut hasher = DefaultHasher::new();
        self.0.hash(&mut hasher);
        hasher.finish()
    }
}

Note: When implementing __hash__ and comparisons, it is important that the following property holds:

k1 == k2 -> hash(k1) == hash(k2)

In other words, if two keys are equal, their hashes must also be equal. In addition you must take care that your classes' hash doesn't change during its lifetime. In this tutorial we do that by not letting Python code change our Number class. In other words, it is immutable.

By default, all #[pyclass] types have a default hash implementation from Python. Types which should not be hashable can override this by setting __hash__ to None. This is the same mechanism as for a pure-Python class. This is done like so:

use pyo3::prelude::*;
#[pyclass]
struct NotHashable {}

#[pymethods]
impl NotHashable {
    #[classattr]
    const __hash__: Option<Py<PyAny>> = None;
}

Comparisons

PyO3 supports the usual magic comparison methods available in Python such as __eq__, __lt__ and so on. It is also possible to support all six operations at once with __richcmp__. This method will be called with a value of CompareOp depending on the operation.

use pyo3::class::basic::CompareOp;

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __richcmp__(&self, other: &Self, op: CompareOp) -> PyResult<bool> {
        match op {
            CompareOp::Lt => Ok(self.0 < other.0),
            CompareOp::Le => Ok(self.0 <= other.0),
            CompareOp::Eq => Ok(self.0 == other.0),
            CompareOp::Ne => Ok(self.0 != other.0),
            CompareOp::Gt => Ok(self.0 > other.0),
            CompareOp::Ge => Ok(self.0 >= other.0),
        }
    }
}

If you obtain the result by comparing two Rust values, as in this example, you can take a shortcut using CompareOp::matches:

use pyo3::class::basic::CompareOp;

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __richcmp__(&self, other: &Self, op: CompareOp) -> bool {
        op.matches(self.0.cmp(&other.0))
    }
}

It checks that the std::cmp::Ordering obtained from Rust's Ord matches the given CompareOp.

Alternatively, you can implement just equality using __eq__:

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __eq__(&self, other: &Self) -> bool {
        self.0 == other.0
    }
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let x = &Bound::new(py, Number(4))?;
        let y = &Bound::new(py, Number(4))?;
        assert!(x.eq(y)?);
        assert!(!x.ne(y)?);
        Ok(())
    })
}

Truthyness

We'll consider Number to be True if it is nonzero:

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __bool__(&self) -> bool {
        self.0 != 0
    }
}

Final code

use std::collections::hash_map::DefaultHasher;
use std::hash::{Hash, Hasher};

use pyo3::prelude::*;
use pyo3::class::basic::CompareOp;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    #[new]
    fn new(value: i32) -> Self {
        Self(value)
    }

    fn __repr__(slf: &Bound<'_, Self>) -> PyResult<String> {
        let class_name: String = slf.get_type().qualname()?;
        Ok(format!("{}({})", class_name, slf.borrow().0))
    }

    fn __str__(&self) -> String {
        self.0.to_string()
    }

    fn __hash__(&self) -> u64 {
        let mut hasher = DefaultHasher::new();
        self.0.hash(&mut hasher);
        hasher.finish()
    }

    fn __richcmp__(&self, other: &Self, op: CompareOp) -> PyResult<bool> {
        match op {
            CompareOp::Lt => Ok(self.0 < other.0),
            CompareOp::Le => Ok(self.0 <= other.0),
            CompareOp::Eq => Ok(self.0 == other.0),
            CompareOp::Ne => Ok(self.0 != other.0),
            CompareOp::Gt => Ok(self.0 > other.0),
            CompareOp::Ge => Ok(self.0 >= other.0),
        }
    }

    fn __bool__(&self) -> bool {
        self.0 != 0
    }
}

#[pymodule]
fn my_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_class::<Number>()?;
    Ok(())
}

Emulating numeric types

At this point we have a Number class that we can't actually do any math on!

Before proceeding, we should think about how we want to handle overflows. There are three obvious solutions:

• We can have infinite precision just like Python's int. However that would be quite boring - we'd be reinventing the wheel.
• We can raise exceptions whenever Number overflows, but that makes the API painful to use.
• We can wrap around the boundary of i32. This is the approach we'll take here. To do that we'll just forward to i32's wrapping_* methods.

Fixing our constructor

Let's address the first overflow, in Number's constructor:

from my_module import Number

n = Number(1 << 1337)

Traceback (most recent call last):
  File "example.py", line 3, in <module>
    n = Number(1 << 1337)
OverflowError: Python int too large to convert to C long

Instead of relying on the default FromPyObject extraction to parse arguments, we can specify our own extraction function, using the #[pyo3(from_py_with = "...")] attribute. Unfortunately PyO3 doesn't provide a way to wrap Python integers out of the box, but we can do a Python call to mask it and cast it to an i32.

#![allow(dead_code)]
use pyo3::prelude::*;

fn wrap(obj: &Bound<'_, PyAny>) -> PyResult<i32> {
    let val = obj.call_method1("__and__", (0xFFFFFFFF_u32,))?;
    let val: u32 = val.extract()?;
    //     👇 This intentionally overflows!
    Ok(val as i32)
}

We also add documentation, via /// comments, which are visible to Python users.

#![allow(dead_code)]
use pyo3::prelude::*;

fn wrap(obj: &Bound<'_, PyAny>) -> PyResult<i32> {
    let val = obj.call_method1("__and__", (0xFFFFFFFF_u32,))?;
    let val: u32 = val.extract()?;
    Ok(val as i32)
}

/// Did you ever hear the tragedy of Darth Signed The Overfloweth? I thought not.
/// It's not a story C would tell you. It's a Rust legend.
#[pyclass(module = "my_module")]
struct Number(i32);

#[pymethods]
impl Number {
    #[new]
    fn new(#[pyo3(from_py_with = "wrap")] value: i32) -> Self {
        Self(value)
    }
}

With that out of the way, let's implement some operators:

use pyo3::exceptions::{PyZeroDivisionError, PyValueError};

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __add__(&self, other: &Self) -> Self {
        Self(self.0.wrapping_add(other.0))
    }

    fn __sub__(&self, other: &Self) -> Self {
        Self(self.0.wrapping_sub(other.0))
    }

    fn __mul__(&self, other: &Self) -> Self {
        Self(self.0.wrapping_mul(other.0))
    }

    fn __truediv__(&self, other: &Self) -> PyResult<Self> {
        match self.0.checked_div(other.0) {
            Some(i) => Ok(Self(i)),
            None => Err(PyZeroDivisionError::new_err("division by zero")),
        }
    }

    fn __floordiv__(&self, other: &Self) -> PyResult<Self> {
        match self.0.checked_div(other.0) {
            Some(i) => Ok(Self(i)),
            None => Err(PyZeroDivisionError::new_err("division by zero")),
        }
    }

    fn __rshift__(&self, other: &Self) -> PyResult<Self> {
        match other.0.try_into() {
            Ok(rhs) => Ok(Self(self.0.wrapping_shr(rhs))),
            Err(_) => Err(PyValueError::new_err("negative shift count")),
        }
    }

    fn __lshift__(&self, other: &Self) -> PyResult<Self> {
        match other.0.try_into() {
            Ok(rhs) => Ok(Self(self.0.wrapping_shl(rhs))),
            Err(_) => Err(PyValueError::new_err("negative shift count")),
        }
    }
}

Unary arithmetic operations

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __pos__(slf: PyRef<'_, Self>) -> PyRef<'_, Self> {
        slf
    }

    fn __neg__(&self) -> Self {
        Self(-self.0)
    }

    fn __abs__(&self) -> Self {
        Self(self.0.abs())
    }

    fn __invert__(&self) -> Self {
        Self(!self.0)
    }
}

Support for the complex(), int() and float() built-in functions.

use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

use pyo3::types::PyComplex;

#[pymethods]
impl Number {
    fn __int__(&self) -> i32 {
        self.0
    }

    fn __float__(&self) -> f64 {
        self.0 as f64
    }

    fn __complex__<'py>(&self, py: Python<'py>) -> Bound<'py, PyComplex> {
        PyComplex::from_doubles_bound(py, self.0 as f64, 0.0)
    }
}

We do not implement the in-place operations like __iadd__ because we do not wish to mutate Number. Similarly we're not interested in supporting operations with different types, so we do not implement the reflected operations like __radd__ either.

Now Python can use our Number class:

from my_module import Number

def hash_djb2(s: str):
	'''
	A version of Daniel J. Bernstein's djb2 string hashing algorithm
	Like many hashing algorithms, it relies on integer wrapping.
	'''

	n = Number(0)
	five = Number(5)

	for x in s:
		n = Number(ord(x)) + ((n << five) - n)
	return n

assert hash_djb2('l50_50') == Number(-1152549421)

Final code

use std::collections::hash_map::DefaultHasher;
use std::hash::{Hash, Hasher};

use pyo3::exceptions::{PyValueError, PyZeroDivisionError};
use pyo3::prelude::*;
use pyo3::class::basic::CompareOp;
use pyo3::types::PyComplex;

fn wrap(obj: &Bound<'_, PyAny>) -> PyResult<i32> {
    let val = obj.call_method1("__and__", (0xFFFFFFFF_u32,))?;
    let val: u32 = val.extract()?;
    Ok(val as i32)
}
/// Did you ever hear the tragedy of Darth Signed The Overfloweth? I thought not.
/// It's not a story C would tell you. It's a Rust legend.
#[pyclass(module = "my_module")]
struct Number(i32);

#[pymethods]
impl Number {
    #[new]
    fn new(#[pyo3(from_py_with = "wrap")] value: i32) -> Self {
        Self(value)
    }

    fn __repr__(slf: &Bound<'_, Self>) -> PyResult<String> {
       // Get the class name dynamically in case `Number` is subclassed
       let class_name: String = slf.get_type().qualname()?;
        Ok(format!("{}({})", class_name, slf.borrow().0))
    }

    fn __str__(&self) -> String {
        self.0.to_string()
    }

    fn __hash__(&self) -> u64 {
        let mut hasher = DefaultHasher::new();
        self.0.hash(&mut hasher);
        hasher.finish()
    }

    fn __richcmp__(&self, other: &Self, op: CompareOp) -> PyResult<bool> {
        match op {
            CompareOp::Lt => Ok(self.0 < other.0),
            CompareOp::Le => Ok(self.0 <= other.0),
            CompareOp::Eq => Ok(self.0 == other.0),
            CompareOp::Ne => Ok(self.0 != other.0),
            CompareOp::Gt => Ok(self.0 > other.0),
            CompareOp::Ge => Ok(self.0 >= other.0),
        }
    }

    fn __bool__(&self) -> bool {
        self.0 != 0
    }

    fn __add__(&self, other: &Self) -> Self {
        Self(self.0.wrapping_add(other.0))
    }

    fn __sub__(&self, other: &Self) -> Self {
        Self(self.0.wrapping_sub(other.0))
    }

    fn __mul__(&self, other: &Self) -> Self {
        Self(self.0.wrapping_mul(other.0))
    }

    fn __truediv__(&self, other: &Self) -> PyResult<Self> {
        match self.0.checked_div(other.0) {
            Some(i) => Ok(Self(i)),
            None => Err(PyZeroDivisionError::new_err("division by zero")),
        }
    }

    fn __floordiv__(&self, other: &Self) -> PyResult<Self> {
        match self.0.checked_div(other.0) {
            Some(i) => Ok(Self(i)),
            None => Err(PyZeroDivisionError::new_err("division by zero")),
        }
    }

    fn __rshift__(&self, other: &Self) -> PyResult<Self> {
        match other.0.try_into() {
            Ok(rhs) => Ok(Self(self.0.wrapping_shr(rhs))),
            Err(_) => Err(PyValueError::new_err("negative shift count")),
        }
    }

    fn __lshift__(&self, other: &Self) -> PyResult<Self> {
        match other.0.try_into() {
            Ok(rhs) => Ok(Self(self.0.wrapping_shl(rhs))),
            Err(_) => Err(PyValueError::new_err("negative shift count")),
        }
    }

    fn __xor__(&self, other: &Self) -> Self {
        Self(self.0 ^ other.0)
    }

    fn __or__(&self, other: &Self) -> Self {
        Self(self.0 | other.0)
    }

    fn __and__(&self, other: &Self) -> Self {
        Self(self.0 & other.0)
    }

    fn __int__(&self) -> i32 {
        self.0
    }

    fn __float__(&self) -> f64 {
        self.0 as f64
    }

    fn __complex__<'py>(&self, py: Python<'py>) -> Bound<'py, PyComplex> {
        PyComplex::from_doubles_bound(py, self.0 as f64, 0.0)
    }
}

#[pymodule]
fn my_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_class::<Number>()?;
    Ok(())
}
const SCRIPT: &'static str = r#"
def hash_djb2(s: str):
    n = Number(0)
    five = Number(5)

    for x in s:
        n = Number(ord(x)) + ((n << five) - n)
    return n

assert hash_djb2('l50_50') == Number(-1152549421)
assert hash_djb2('logo') == Number(3327403)
assert hash_djb2('horizon') == Number(1097468315)


assert Number(2) + Number(2) == Number(4)
assert Number(2) + Number(2) != Number(5)

assert Number(13) - Number(7) == Number(6)
assert Number(13) - Number(-7) == Number(20)

assert Number(13) / Number(7) == Number(1)
assert Number(13) // Number(7) == Number(1)

assert Number(13) * Number(7) == Number(13*7)

assert Number(13) > Number(7)
assert Number(13) < Number(20)
assert Number(13) == Number(13)
assert Number(13) >= Number(7)
assert Number(13) <= Number(20)
assert Number(13) == Number(13)


assert (True if Number(1) else False)
assert (False if Number(0) else True)


assert int(Number(13)) == 13
assert float(Number(13)) == 13
assert Number.__doc__ == "Did you ever hear the tragedy of Darth Signed The Overfloweth? I thought not.\nIt's not a story C would tell you. It's a Rust legend."
assert Number(12345234523452) == Number(1498514748)
try:
    import inspect
    assert inspect.signature(Number).__str__() == '(value)'
except ValueError:
    # Not supported with `abi3` before Python 3.10
    pass
assert Number(1337).__str__() == '1337'
assert Number(1337).__repr__() == 'Number(1337)'
"#;


use pyo3::PyTypeInfo;

fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        let globals = PyModule::import_bound(py, "__main__")?.dict();
        globals.set_item("Number", Number::type_object_bound(py))?;

        py.run_bound(SCRIPT, Some(&globals), None)?;
        Ok(())
    })
}

Appendix: Writing some unsafe code

At the beginning of this chapter we said that PyO3 doesn't provide a way to wrap Python integers out of the box but that's a half truth. There's not a PyO3 API for it, but there's a Python C API function that does:

unsigned long PyLong_AsUnsignedLongMask(PyObject *obj)

We can call this function from Rust by using pyo3::ffi::PyLong_AsUnsignedLongMask. This is an unsafe function, which means we have to use an unsafe block to call it and take responsibility for upholding the contracts of this function. Let's review those contracts:

• The GIL must be held. If it's not, calling this function causes a data race.
• The pointer must be valid, i.e. it must be properly aligned and point to a valid Python object.

Let's create that helper function. The signature has to be fn(&Bound<'_, PyAny>) -> PyResult<T>.

• &Bound<'_, PyAny> represents a checked borrowed reference, so the pointer derived from it is valid (and not null).
• Whenever we have borrowed references to Python objects in scope, it is guaranteed that the GIL is held. This reference is also where we can get a Python token to use in our call to PyErr::take.

#![allow(dead_code)]
use std::os::raw::c_ulong;
use pyo3::prelude::*;
use pyo3::ffi;

fn wrap(obj: &Bound<'_, PyAny>) -> Result<i32, PyErr> {
    let py: Python<'_> = obj.py();

    unsafe {
        let ptr = obj.as_ptr();

        let ret: c_ulong = ffi::PyLong_AsUnsignedLongMask(ptr);
        if ret == c_ulong::MAX {
            if let Some(err) = PyErr::take(py) {
                return Err(err);
            }
        }

        Ok(ret as i32)
    }
}

Emulating callable objects

Classes can be callable if they have a #[pymethod] named __call__. This allows instances of a class to behave similar to functions.

This method's signature must look like __call__(<self>, ...) -> object - here, any argument list can be defined as for normal pymethods

Example: Implementing a call counter

The following pyclass is a basic decorator - its constructor takes a Python object as argument and calls that object when called. An equivalent Python implementation is linked at the end.

An example crate containing this pyclass can be found here

use pyo3::prelude::*;
use pyo3::types::{PyDict, PyTuple};
use std::cell::Cell;

/// A function decorator that keeps track how often it is called.
///
/// It otherwise doesn't do anything special.
#[pyclass(name = "Counter")]
pub struct PyCounter {
    // Keeps track of how many calls have gone through.
    //
    // See the discussion at the end for why `Cell` is used.
    count: Cell<u64>,

    // This is the actual function being wrapped.
    wraps: Py<PyAny>,
}

#[pymethods]
impl PyCounter {
    // Note that we don't validate whether `wraps` is actually callable.
    //
    // While we could use `PyAny::is_callable` for that, it has some flaws:
    //    1. It doesn't guarantee the object can actually be called successfully
    //    2. We still need to handle any exceptions that the function might raise
    #[new]
    fn __new__(wraps: Py<PyAny>) -> Self {
        PyCounter {
            count: Cell::new(0),
            wraps,
        }
    }

    #[getter]
    fn count(&self) -> u64 {
        self.count.get()
    }

    #[pyo3(signature = (*args, **kwargs))]
    fn __call__(
        &self,
        py: Python<'_>,
        args: &Bound<'_, PyTuple>,
        kwargs: Option<&Bound<'_, PyDict>>,
    ) -> PyResult<Py<PyAny>> {
        let old_count = self.count.get();
        let new_count = old_count + 1;
        self.count.set(new_count);
        let name = self.wraps.getattr(py, "__name__")?;

        println!("{} has been called {} time(s).", name, new_count);

        // After doing something, we finally forward the call to the wrapped function
        let ret = self.wraps.call_bound(py, args, kwargs)?;

        // We could do something with the return value of
        // the function before returning it
        Ok(ret)
    }
}

#[pymodule]
pub fn decorator(module: &Bound<'_, PyModule>) -> PyResult<()> {
    module.add_class::<PyCounter>()?;
    Ok(())
}

Python code:

from decorator import Counter


@Counter
def say_hello():
    print("hello")


say_hello()
say_hello()
say_hello()
say_hello()

assert say_hello.count == 4

Output:

say_hello has been called 1 time(s).
hello
say_hello has been called 2 time(s).
hello
say_hello has been called 3 time(s).
hello
say_hello has been called 4 time(s).
hello

Pure Python implementation

A Python implementation of this looks similar to the Rust version:

class Counter:
    def __init__(self, wraps):
        self.count = 0
        self.wraps = wraps

    def __call__(self, *args, **kwargs):
        self.count += 1
        print(f"{self.wraps.__name__} has been called {self.count} time(s)")
        self.wraps(*args, **kwargs)

Note that it can also be implemented as a higher order function:

def Counter(wraps):
    count = 0
    def call(*args, **kwargs):
        nonlocal count
        count += 1
        print(f"{wraps.__name__} has been called {count} time(s)")
        return wraps(*args, **kwargs)
    return call

What is the Cell for?

A previous implementation used a normal u64, which meant it required a &mut self receiver to update the count:

#[pyo3(signature = (*args, **kwargs))]
fn __call__(
    &mut self,
    py: Python<'_>,
    args: &Bound<'_, PyTuple>,
    kwargs: Option<&Bound<'_, PyDict>>,
) -> PyResult<Py<PyAny>> {
    self.count += 1;
    let name = self.wraps.getattr(py, "__name__")?;

    println!("{} has been called {} time(s).", name, self.count);

    // After doing something, we finally forward the call to the wrapped function
    let ret = self.wraps.call(py, args, kwargs)?;

    // We could do something with the return value of
    // the function before returning it
    Ok(ret)
}

The problem with this is that the &mut self receiver means PyO3 has to borrow it exclusively, and hold this borrow across theself.wraps.call(py, args, kwargs) call. This call returns control to the user's Python code which is free to call arbitrary things, including the decorated function. If that happens PyO3 is unable to create a second unique borrow and will be forced to raise an exception.

As a result, something innocent like this will raise an exception:

@Counter
def say_hello():
    if say_hello.count < 2:
        print(f"hello from decorator")

say_hello()
# RuntimeError: Already borrowed

The implementation in this chapter fixes that by never borrowing exclusively; all the methods take &self as receivers, of which multiple may exist simultaneously. This requires a shared counter and the easiest way to do that is to use Cell, so that's what is used here.

This shows the dangers of running arbitrary Python code - note that "running arbitrary Python code" can be far more subtle than the example above:

• Python's asynchronous executor may park the current thread in the middle of Python code, even in Python code that you control, and let other Python code run.
• Dropping arbitrary Python objects may invoke destructors defined in Python (__del__ methods).
• Calling Python's C-api (most PyO3 apis call C-api functions internally) may raise exceptions, which may allow Python code in signal handlers to run.

This is especially important if you are writing unsafe code; Python code must never be able to cause undefined behavior. You must ensure that your Rust code is in a consistent state before doing any of the above things.

Calling Python in Rust code

This chapter of the guide documents some ways to interact with Python code from Rust.

Below is an introduction to the 'py lifetime and some general remarks about how PyO3's API reasons about Python code.

The subchapters also cover the following topics:

• Python object types available in PyO3's API
• How to work with Python exceptions
• How to call Python functions
• How to execute existing Python code

The 'py lifetime

To safely interact with the Python interpreter a Rust thread must have a corresponding Python thread state and hold the Global Interpreter Lock (GIL). PyO3 has a Python<'py> token that is used to prove that these conditions are met. Its lifetime 'py is a central part of PyO3's API.

The Python<'py> token serves three purposes:

• It provides global APIs for the Python interpreter, such as [py.eval_bound()][eval] and [py.import_bound()][import].
• It can be passed to functions that require a proof of holding the GIL, such as [Py::clone_ref][clone_ref].
• Its lifetime 'py is used to bind many of PyO3's types to the Python interpreter, such as [Bound<'py, T>][Bound].

PyO3's types that are bound to the 'py lifetime, for example Bound<'py, T>, all contain a Python<'py> token. This means they have full access to the Python interpreter and offer a complete API for interacting with Python objects.

Consult PyO3's API documentation to learn how to acquire one of these tokens.

The Global Interpreter Lock

Concurrent programming in Python is aided by the Global Interpreter Lock (GIL), which ensures that only one Python thread can use the Python interpreter and its API at the same time. This allows it to be used to synchronize code. See the pyo3::sync module for synchronization tools PyO3 offers that are based on the GIL's guarantees.

Non-Python operations (system calls and native Rust code) can unlock the GIL. See the section on parallelism for how to do that using PyO3's API.

Python's memory model

Python's memory model differs from Rust's memory model in two key ways:

• There is no concept of ownership; all Python objects are shared and usually implemented via reference counting
• There is no concept of exclusive (&mut) references; any reference can mutate a Python object

PyO3's API reflects this by providing smart pointer types, Py<T>, Bound<'py, T>, and (the very rarely used) Borrowed<'a, 'py, T>. These smart pointers all use Python reference counting. See the subchapter on types for more detail on these types.

Because of the lack of exclusive &mut references, PyO3's APIs for Python objects, for example [PyListMethods::append], use shared references. This is safe because Python objects have internal mechanisms to prevent data races (as of time of writing, the Python GIL).

Python object types

PyO3 offers two main sets of types to interact with Python objects. This section of the guide expands into detail about these types and how to choose which to use.

The first set of types is are the smart pointers which all Python objects are wrapped in. These are Py<T>, Bound<'py, T>, and Borrowed<'a, 'py, T>. The first section below expands on each of these in detail and why there are three of them.

The second set of types are types which fill in the generic parameter T of the smart pointers. The most common is PyAny, which represents any Python object (similar to Python's typing.Any). There are also concrete types for many Python built-in types, such as PyList, PyDict, and PyTuple. User defined #[pyclass] types also fit this category. The second section below expands on how to use these types.

Before PyO3 0.21, PyO3's main API to interact with Python objects was a deprecated API known as the "GIL Refs" API, containing reference types such as &PyAny, &PyList, and &PyCell<T> for user-defined #[pyclass] types. The third section below details this deprecated API.

PyO3's smart pointers

PyO3's API offers three generic smart pointers: Py<T>, Bound<'py, T> and Borrowed<'a, 'py, T>. For each of these the type parameter T will be filled by a concrete Python type. For example, a Python list object can be represented by Py<PyList>, Bound<'py, PyList>, and Borrowed<'a, 'py, PyList>.

These smart pointers behave differently due to their lifetime parameters. Py<T> has no lifetime parameters, Bound<'py, T> has the 'py lifetime as a parameter, and Borrowed<'a, 'py, T> has the 'py lifetime plus an additional lifetime 'a to denote the lifetime it is borrowing data for. (You can read more about these lifetimes in the subsections below).

Python objects are reference counted, like std::sync::Arc. A major reason for these smart pointers is to bring Python's reference counting to a Rust API.

The recommendation of when to use each of these smart pointers is as follows:

• Use Bound<'py, T> for as much as possible, as it offers the most efficient and complete API.
• Use Py<T> mostly just for storage inside Rust structs which do not want to or can't add a lifetime parameter for Bound<'py, T>.
• Borrowed<'a, 'py, T> is almost never used. It is occasionally present at the boundary between Rust and the Python interpreter, for example when borrowing data from Python tuples (which is safe because they are immutable).

The sections below also explain these smart pointers in a little more detail.

Py<T> (and PyObject)

Py<T> is the foundational smart pointer in PyO3's API. The type parameter T denotes the type of the Python object. Very frequently this is PyAny, meaning any Python object. This is so common that Py<PyAny> has a type alias PyObject.

Because Py<T> is not bound to the 'py lifetime, it is the type to use when storing a Python object inside a Rust struct or enum which do not want to have a lifetime parameter. In particular, #[pyclass] types are not permitted to have a lifetime, so Py<T> is the correct type to store Python objects inside them.

The lack of binding to the 'py lifetime also carries drawbacks:

• Almost all methods on Py<T> require a Python<'py> token as the first argument
• Other functionality, such as Drop, needs to check at runtime for attachment to the Python GIL, at a small performance cost

Because of the drawbacks Bound<'py, T> is preferred for many of PyO3's APIs. In particular, Bound<'py, T> is the better for function arguments.

To convert a Py<T> into a Bound<'py, T>, the Py::bind and Py::into_bound methods are available. Bound<'py, T> can be converted back into Py<T> using Bound::unbind.

Bound<'py, T>

Bound<'py, T> is the counterpart to Py<T> which is also bound to the 'py lifetime. It can be thought of as equivalent to the Rust tuple (Python<'py>, Py<T>).

By having the binding to the 'py lifetime, Bound<'py, T> can offer the complete PyO3 API at maximum efficiency. This means that in almost all cases where Py<T> is not necessary for lifetime reasons, Bound<'py, T> should be used.

Bound<'py, T> engages in Python reference counting. This means that Bound<'py, T> owns a Python object. Rust code which just wants to borrow a Python object should use a shared reference &Bound<'py, T>. Just like std::sync::Arc, using .clone() and drop() will cheaply increment and decrement the reference count of the object (just in this case, the reference counting is implemented by the Python interpreter itself).

To give an example of how Bound<'py, T> is PyO3's primary API type, consider the following Python code:

def example():
    x = list()   # create a Python list
    x.append(1)  # append the integer 1 to it
    y = x        # create a second reference to the list
    del x        # delete the original reference

Using PyO3's API, and in particular Bound<'py, PyList>, this code translates into the following Rust code:

use pyo3::prelude::*;
use pyo3::types::PyList;

fn example<'py>(py: Python<'py>) -> PyResult<()> {
    let x: Bound<'py, PyList> = PyList::empty_bound(py);
    x.append(1)?;
    let y: Bound<'py, PyList> = x.clone(); // y is a new reference to the same list
    drop(x); // release the original reference x
    Ok(())
}
Python::with_gil(example).unwrap();

Or, without the type annotations:

use pyo3::prelude::*;
use pyo3::types::PyList;

fn example(py: Python<'_>) -> PyResult<()> {
    let x = PyList::empty_bound(py);
    x.append(1)?;
    let y = x.clone();
    drop(x);
    Ok(())
}
Python::with_gil(example).unwrap();

Function argument lifetimes

Because the 'py lifetime often appears in many function arguments as part of the Bound<'py, T> smart pointer, the Rust compiler will often require annotations of input and output lifetimes. This occurs when the function output has at least one lifetime, and there is more than one lifetime present on the inputs.

To demonstrate, consider this function which takes accepts Python objects and applies the Python + operation to them:

use pyo3::prelude::*;
fn add(left: &'_ Bound<'_, PyAny>, right: &'_ Bound<'_, PyAny>) -> PyResult<Bound<'_, PyAny>> {
    left.add(right)
}

Because the Python + operation might raise an exception, this function returns PyResult<Bound<'_, PyAny>>. It doesn't need ownership of the inputs, so it takes &Bound<'_, PyAny> shared references. To demonstrate the point, all lifetimes have used the wildcard '_ to allow the Rust compiler to attempt to infer them. Because there are four input lifetimes (two lifetimes of the shared references, and two 'py lifetimes unnamed inside the Bound<'_, PyAny> pointers), the compiler cannot reason about which must be connected to the output.

The correct way to solve this is to add the 'py lifetime as a parameter for the function, and name all the 'py lifetimes inside the Bound<'py, PyAny> smart pointers. For the shared references, it's also fine to reduce &'_ to just &. The working end result is below:

use pyo3::prelude::*;
fn add<'py>(
    left: &Bound<'py, PyAny>,
    right: &Bound<'py, PyAny>,
) -> PyResult<Bound<'py, PyAny>> {
    left.add(right)
}
Python::with_gil(|py| {
    let s = pyo3::types::PyString::new_bound(py, "s");
    assert!(add(&s, &s).unwrap().eq("ss").unwrap());
})

If naming the 'py lifetime adds unwanted complexity to the function signature, it is also acceptable to return PyObject (aka Py<PyAny>), which has no lifetime. The cost is instead paid by a slight increase in implementation complexity, as seen by the introduction of a call to Bound::unbind:

use pyo3::prelude::*;
fn add(left: &Bound<'_, PyAny>, right: &Bound<'_, PyAny>) -> PyResult<PyObject> {
    let output: Bound<'_, PyAny> = left.add(right)?;
    Ok(output.unbind())
}
Python::with_gil(|py| {
    let s = pyo3::types::PyString::new_bound(py, "s");
    assert!(add(&s, &s).unwrap().bind(py).eq("ss").unwrap());
})

Borrowed<'a, 'py, T>

Borrowed<'a, 'py, T> is an advanced type used just occasionally at the edge of interaction with the Python interpreter. It can be thought of as analogous to the shared reference &'a Bound<'py, T>. The difference is that Borrowed<'a, 'py, T> is just a smart pointer rather than a reference-to-a-smart-pointer, which is a helpful reduction in indirection in specific interactions with the Python interpreter.

Borrowed<'a, 'py, T> dereferences to Bound<'py, T>, so all methods on Bound<'py, T> are available on Borrowed<'a, 'py, T>.

An example where Borrowed<'a, 'py, T> is used is in PyTupleMethods::get_borrowed_item:

use pyo3::prelude::*;
use pyo3::types::PyTuple;

fn example<'py>(py: Python<'py>) -> PyResult<()> {
// Create a new tuple with the elements (0, 1, 2)
let t = PyTuple::new_bound(py, [0, 1, 2]);
for i in 0..=2 {
    let entry: Borrowed<'_, 'py, PyAny> = t.get_borrowed_item(i)?;
    // `PyAnyMethods::extract` is available on `Borrowed`
    // via the dereference to `Bound`
    let value: usize = entry.extract()?;
    assert_eq!(i, value);
}
Ok(())
}
Python::with_gil(example).unwrap();

Casting between smart pointer types

To convert between Py<T> and Bound<'py, T> use the bind() / into_bound() methods. Use the as_unbound() / unbind() methods to go back from Bound<'py, T> to Py<T>.

let obj: Py<PyAny> = ...;
let bound: &Bound<'py, PyAny> = obj.bind(py);
let bound: Bound<'py, PyAny> = obj.into_bound(py);

let obj: &Py<PyAny> = bound.as_unbound();
let obj: Py<PyAny> = bound.unbind();

To convert between Bound<'py, T> and Borrowed<'a, 'py, T> use the as_borrowed() method. Borrowed<'a, 'py, T> has a deref coercion to Bound<'py, T>. Use the to_owned() method to increment the Python reference count and to create a new Bound<'py, T> from the Borrowed<'a, 'py, T>.

let bound: Bound<'py, PyAny> = ...;
let borrowed: Borrowed<'_, 'py, PyAny> = bound.as_borrowed();

// deref coercion
let bound: &Bound<'py, PyAny> = &borrowed;

// create a new Bound by increase the Python reference count
let bound: Bound<'py, PyAny> = borrowed.to_owned();

To convert between Py<T> and Borrowed<'a, 'py, T> use the bind_borrowed() method. Use either as_unbound() or .to_owned().unbind() to go back to Py<T> from Borrowed<'a, 'py, T>, via Bound<'py, T>.

let obj: Py<PyAny> = ...;
let borrowed: Borrowed<'_, 'py, PyAny> = bound.as_borrowed();

// via deref coercion to Bound and then using Bound::as_unbound
let obj: &Py<PyAny> = borrowed.as_unbound();

// via a new Bound by increasing the Python reference count, and unbind it
let obj: Py<PyAny> = borrowed.to_owned().unbind().

Concrete Python types

In all of Py<T>, Bound<'py, T>, and Borrowed<'a, 'py, T>, the type parameter T denotes the type of the Python object referred to by the smart pointer.

This parameter T can be filled by:

• PyAny, which represents any Python object,
• Native Python types such as PyList, PyTuple, and PyDict, and
• #[pyclass] types defined from Rust

The following subsections covers some further detail about how to work with these types:

• the APIs that are available for these concrete types,
• how to cast Bound<'py, T> to a specific concrete type, and
• how to get Rust data out of a Bound<'py, T>.

Using APIs for concrete Python types

Each concrete Python type such as PyAny, PyTuple and PyDict exposes its API on the corresponding bound smart pointer Bound<'py, PyAny>, Bound<'py, PyTuple> and Bound<'py, PyDict>.

Each type's API is exposed as a trait: PyAnyMethods, PyTupleMethods, PyDictMethods, and so on for all concrete types. Using traits rather than associated methods on the Bound smart pointer is done for a couple of reasons:

• Clarity of documentation: each trait gets its own documentation page in the PyO3 API docs. If all methods were on the Bound smart pointer directly, the vast majority of PyO3's API would be on a single, extremely long, documentation page.
• Consistency: downstream code implementing Rust APIs for existing Python types can also follow this pattern of using a trait. Downstream code would not be allowed to add new associated methods directly on the Bound type.
• Future design: it is hoped that a future Rust with arbitrary self types will remove the need for these traits in favour of placing the methods directly on PyAny, PyTuple, PyDict, and so on.

These traits are all included in the pyo3::prelude module, so with the glob import use pyo3::prelude::* the full PyO3 API is made available to downstream code.

The following function accesses the first item in the input Python list, using the .get_item() method from the PyListMethods trait:

use pyo3::prelude::*;
use pyo3::types::PyList;

fn get_first_item<'py>(list: &Bound<'py, PyList>) -> PyResult<Bound<'py, PyAny>> {
    list.get_item(0)
}
Python::with_gil(|py| {
    let l = PyList::new_bound(py, ["hello world"]);
    assert!(get_first_item(&l).unwrap().eq("hello world").unwrap());
})

Casting between Python object types

To cast Bound<'py, T> smart pointers to some other type, use the .downcast() family of functions. This converts &Bound<'py, T> to a different &Bound<'py, U>, without transferring ownership. There is also .downcast_into() to convert Bound<'py, T> to Bound<'py, U> with transfer of ownership. These methods are available for all types T which implement the PyTypeCheck trait.

Casting to Bound<'py, PyAny> can be done with .as_any() or .into_any().

For example, the following snippet shows how to cast Bound<'py, PyAny> to Bound<'py, PyTuple>:

use pyo3::prelude::*;
use pyo3::types::PyTuple;
fn example<'py>(py: Python<'py>) -> PyResult<()> {
// create a new Python `tuple`, and use `.into_any()` to erase the type
let obj: Bound<'py, PyAny> = PyTuple::empty_bound(py).into_any();

// use `.downcast()` to cast to `PyTuple` without transferring ownership
let _: &Bound<'py, PyTuple> = obj.downcast()?;

// use `.downcast_into()` to cast to `PyTuple` with transfer of ownership
let _: Bound<'py, PyTuple> = obj.downcast_into()?;
Ok(())
}
Python::with_gil(example).unwrap()

Custom #[pyclass] types implement PyTypeCheck, so .downcast() also works for these types. The snippet below is the same as the snippet above casting instead to a custom type MyClass:

use pyo3::prelude::*;

#[pyclass]
struct MyClass {}

fn example<'py>(py: Python<'py>) -> PyResult<()> {
// create a new Python `tuple`, and use `.into_any()` to erase the type
let obj: Bound<'py, PyAny> = Bound::new(py, MyClass {})?.into_any();

// use `.downcast()` to cast to `MyClass` without transferring ownership
let _: &Bound<'py, MyClass> = obj.downcast()?;

// use `.downcast_into()` to cast to `MyClass` with transfer of ownership
let _: Bound<'py, MyClass> = obj.downcast_into()?;
Ok(())
}
Python::with_gil(example).unwrap()

Extracting Rust data from Python objects

To extract Rust data from Python objects, use .extract() instead of .downcast(). This method is available for all types which implement the [FromPyObject] trait.

For example, the following snippet extracts a Rust tuple of integers from a Python tuple:

use pyo3::prelude::*;
use pyo3::types::PyTuple;
fn example<'py>(py: Python<'py>) -> PyResult<()> {
// create a new Python `tuple`, and use `.into_any()` to erase the type
let obj: Bound<'py, PyAny> = PyTuple::new_bound(py, [1, 2, 3]).into_any();

// extracting the Python `tuple` to a rust `(i32, i32, i32)` tuple
let (x, y, z) = obj.extract::<(i32, i32, i32)>()?;
assert_eq!((x, y, z), (1, 2, 3));
Ok(())
}
Python::with_gil(example).unwrap()

To avoid copying data, #[pyclass] types can directly reference Rust data stored within the Python objects without needing to .extract(). See the corresponding documentation in the class section of the guide for more detail.

The GIL Refs API

The GIL Refs API was PyO3's primary API prior to PyO3 0.21. The main difference was that instead of the Bound<'py, PyAny> smart pointer, the "GIL Reference" &'py PyAny was used. (This was similar for other Python types.)

As of PyO3 0.21, the GIL Refs API is deprecated. See the migration guide for details on how to upgrade.

The following sections note some historical detail about the GIL Refs API.

PyAny

Represented: a Python object of unspecified type. In the GIL Refs API, this was only accessed as the GIL Ref &'py PyAny.

Used: &'py PyAny was used to refer to some Python object when the GIL lifetime was available for the whole duration access was needed. For example, intermediate values and arguments to pyfunctions or pymethods implemented in Rust where any type is allowed.

Conversions:

For a &PyAny object reference any where the underlying object is a Python-native type such as a list:

use pyo3::prelude::*;
use pyo3::types::PyList;
Python::with_gil(|py| -> PyResult<()> {
#[allow(deprecated)] // PyList::empty is part of the deprecated "GIL Refs" API.
let obj: &PyAny = PyList::empty(py);

// To &PyList with PyAny::downcast
let _: &PyList = obj.downcast()?;

// To Py<PyAny> (aka PyObject) with .into()
let _: Py<PyAny> = obj.into();

// To Py<PyList> with PyAny::extract
let _: Py<PyList> = obj.extract()?;
Ok(())
}).unwrap();

For a &PyAny object reference any where the underlying object is a #[pyclass]:

use pyo3::prelude::*;
#[pyclass] #[derive(Clone)] struct MyClass { }
Python::with_gil(|py| -> PyResult<()> {
#[allow(deprecated)] // into_ref is part of the deprecated GIL Refs API
let obj: &PyAny = Py::new(py, MyClass {})?.into_ref(py);

// To &PyCell<MyClass> with PyAny::downcast
#[allow(deprecated)] // &PyCell is part of the deprecated GIL Refs API
let _: &PyCell<MyClass> = obj.downcast()?;

// To Py<PyAny> (aka PyObject) with .into()
let _: Py<PyAny> = obj.into();

// To Py<MyClass> with PyAny::extract
let _: Py<MyClass> = obj.extract()?;

// To MyClass with PyAny::extract, if MyClass: Clone
let _: MyClass = obj.extract()?;

// To PyRef<'_, MyClass> or PyRefMut<'_, MyClass> with PyAny::extract
let _: PyRef<'_, MyClass> = obj.extract()?;
let _: PyRefMut<'_, MyClass> = obj.extract()?;
Ok(())
}).unwrap();

PyTuple, PyDict, and many more

Represented: a native Python object of known type. In the GIL Refs API, they were only accessed as the GIL Refs &'py PyTuple, &'py PyDict.

Used: &'py PyTuple and similar were used to operate with native Python types while holding the GIL. Like PyAny, this is the most convenient form to use for function arguments and intermediate values.

These GIL Refs implement Deref<Target = PyAny>, so they all expose the same methods which can be found on PyAny.

To see all Python types exposed by PyO3 consult the pyo3::types module.

Conversions:

use pyo3::prelude::*;
use pyo3::types::PyList;
Python::with_gil(|py| -> PyResult<()> {
#[allow(deprecated)] // PyList::empty is part of the deprecated "GIL Refs" API.
let list = PyList::empty(py);

// Use methods from PyAny on all Python types with Deref implementation
let _ = list.repr()?;

// To &PyAny automatically with Deref implementation
let _: &PyAny = list;

// To &PyAny explicitly with .as_ref()
#[allow(deprecated)] // as_ref is part of the deprecated "GIL Refs" API.
let _: &PyAny = list.as_ref();

// To Py<T> with .into() or Py::from()
let _: Py<PyList> = list.into();

// To PyObject with .into() or .to_object(py)
let _: PyObject = list.into();
Ok(())
}).unwrap();

Py<T> and PyObject

Represented: a GIL-independent reference to a Python object. This can be a Python native type (like PyTuple), or a pyclass type implemented in Rust. The most commonly-used variant, Py<PyAny>, is also known as PyObject.

Used: Whenever you want to carry around references to a Python object without caring about a GIL lifetime. For example, storing Python object references in a Rust struct that outlives the Python-Rust FFI boundary, or returning objects from functions implemented in Rust back to Python.

Can be cloned using Python reference counts with .clone().

PyCell<SomeType>

Represented: a reference to a Rust object (instance of PyClass) wrapped in a Python object. The cell part is an analog to stdlib's RefCell to allow access to &mut references.

Used: for accessing pure-Rust API of the instance (members and functions taking &SomeType or &mut SomeType) while maintaining the aliasing rules of Rust references.

Like PyO3's Python native types, the GIL Ref &PyCell<T> implements Deref<Target = PyAny>, so it also exposed all of the methods on PyAny.

Conversions:

PyCell<T> was used to access &T and &mut T via PyRef<T> and PyRefMut<T> respectively.

use pyo3::prelude::*;
#[pyclass] struct MyClass { }
Python::with_gil(|py| -> PyResult<()> {
#[allow(deprecated)] // &PyCell is part of the deprecated GIL Refs API
let cell: &PyCell<MyClass> = PyCell::new(py, MyClass {})?;

// To PyRef<T> with .borrow() or .try_borrow()
let py_ref: PyRef<'_, MyClass> = cell.try_borrow()?;
let _: &MyClass = &*py_ref;
drop(py_ref);

// To PyRefMut<T> with .borrow_mut() or .try_borrow_mut()
let mut py_ref_mut: PyRefMut<'_, MyClass> = cell.try_borrow_mut()?;
let _: &mut MyClass = &mut *py_ref_mut;
Ok(())
}).unwrap();

PyCell<T> was also accessed like a Python-native type.

use pyo3::prelude::*;
#[pyclass] struct MyClass { }
Python::with_gil(|py| -> PyResult<()> {
#[allow(deprecated)] // &PyCell is part of the deprecate GIL Refs API
let cell: &PyCell<MyClass> = PyCell::new(py, MyClass {})?;

// Use methods from PyAny on PyCell<T> with Deref implementation
let _ = cell.repr()?;

// To &PyAny automatically with Deref implementation
let _: &PyAny = cell;

// To &PyAny explicitly with .as_ref()
#[allow(deprecated)] // as_ref is part of the deprecated "GIL Refs" API.
let _: &PyAny = cell.as_ref();
Ok(())
}).unwrap();

Python exceptions

Defining a new exception

Use the create_exception! macro:

use pyo3::create_exception;

create_exception!(module, MyError, pyo3::exceptions::PyException);

• module is the name of the containing module.
• MyError is the name of the new exception type.

For example:

use pyo3::prelude::*;
use pyo3::create_exception;
use pyo3::types::IntoPyDict;
use pyo3::exceptions::PyException;

create_exception!(mymodule, CustomError, PyException);

Python::with_gil(|py| {
    let ctx = [("CustomError", py.get_type_bound::<CustomError>())].into_py_dict_bound(py);
    pyo3::py_run!(
        py,
        *ctx,
        "assert str(CustomError) == \"<class 'mymodule.CustomError'>\""
    );
    pyo3::py_run!(py, *ctx, "assert CustomError('oops').args == ('oops',)");
});

When using PyO3 to create an extension module, you can add the new exception to the module like this, so that it is importable from Python:

use pyo3::prelude::*;
use pyo3::exceptions::PyException;

pyo3::create_exception!(mymodule, CustomError, PyException);

#[pymodule]
fn mymodule(py: Python<'_>, m: &Bound<'_, PyModule>) -> PyResult<()> {
    // ... other elements added to module ...
    m.add("CustomError", py.get_type_bound::<CustomError>())?;

    Ok(())
}

Raising an exception

As described in the function error handling chapter, to raise an exception from a #[pyfunction] or #[pymethods], return an Err(PyErr). PyO3 will automatically raise this exception for you when returning the result to Python.

You can also manually write and fetch errors in the Python interpreter's global state:

use pyo3::{Python, PyErr};
use pyo3::exceptions::PyTypeError;

Python::with_gil(|py| {
    PyTypeError::new_err("Error").restore(py);
    assert!(PyErr::occurred(py));
    drop(PyErr::fetch(py));
});

Checking exception types

Python has an isinstance method to check an object's type. In PyO3 every object has the PyAny::is_instance and PyAny::is_instance_of methods which do the same thing.

use pyo3::prelude::*;
use pyo3::types::{PyBool, PyList};

Python::with_gil(|py| {
    assert!(PyBool::new_bound(py, true).is_instance_of::<PyBool>());
    let list = PyList::new_bound(py, &[1, 2, 3, 4]);
    assert!(!list.is_instance_of::<PyBool>());
    assert!(list.is_instance_of::<PyList>());
});

To check the type of an exception, you can similarly do:

use pyo3::exceptions::PyTypeError;
use pyo3::prelude::*;
Python::with_gil(|py| {
let err = PyTypeError::new_err(());
err.is_instance_of::<PyTypeError>(py);
});

Using exceptions defined in Python code

It is possible to use an exception defined in Python code as a native Rust type. The import_exception! macro allows importing a specific exception class and defines a Rust type for that exception.

#![allow(dead_code)]
use pyo3::prelude::*;

mod io {
    pyo3::import_exception!(io, UnsupportedOperation);
}

fn tell(file: &Bound<'_, PyAny>) -> PyResult<u64> {
    match file.call_method0("tell") {
        Err(_) => Err(io::UnsupportedOperation::new_err("not supported: tell")),
        Ok(x) => x.extract::<u64>(),
    }
}

pyo3::exceptions defines exceptions for several standard library modules.

Calling Python functions

The Bound<'py, T> smart pointer (such as Bound<'py, PyAny>, Bound<'py, PyList>, or Bound<'py, MyClass>) can be used to call Python functions.

PyO3 offers two APIs to make function calls:

• call - call any callable Python object.
• call_method - call a method on the Python object.

Both of these APIs take args and kwargs arguments (for positional and keyword arguments respectively). There are variants for less complex calls:

• call1 and call_method1 to call only with positional args.
• call0 and call_method0 to call with no arguments.

For convenience the Py<T> smart pointer also exposes these same six API methods, but needs a Python token as an additional first argument to prove the GIL is held.

The example below calls a Python function behind a PyObject (aka Py<PyAny>) reference:

use pyo3::prelude::*;
use pyo3::types::PyTuple;

fn main() -> PyResult<()> {
    let arg1 = "arg1";
    let arg2 = "arg2";
    let arg3 = "arg3";

    Python::with_gil(|py| {
        let fun: Py<PyAny> = PyModule::from_code_bound(
            py,
            "def example(*args, **kwargs):
                if args != ():
                    print('called with args', args)
                if kwargs != {}:
                    print('called with kwargs', kwargs)
                if args == () and kwargs == {}:
                    print('called with no arguments')",
            "",
            "",
        )?
        .getattr("example")?
        .into();

        // call object without any arguments
        fun.call0(py)?;

        // pass object with Rust tuple of positional arguments
        let args = (arg1, arg2, arg3);
        fun.call1(py, args)?;

        // call object with Python tuple of positional arguments
        let args = PyTuple::new_bound(py, &[arg1, arg2, arg3]);
        fun.call1(py, args)?;
        Ok(())
    })
}

Creating keyword arguments

For the call and call_method APIs, kwargs are Option<&Bound<'py, PyDict>>, so can either be None or Some(&dict). You can use the IntoPyDict trait to convert other dict-like containers, e.g. HashMap or BTreeMap, as well as tuples with up to 10 elements and Vecs where each element is a two-element tuple.

use pyo3::prelude::*;
use pyo3::types::IntoPyDict;
use std::collections::HashMap;

fn main() -> PyResult<()> {
    let key1 = "key1";
    let val1 = 1;
    let key2 = "key2";
    let val2 = 2;

    Python::with_gil(|py| {
        let fun: Py<PyAny> = PyModule::from_code_bound(
            py,
            "def example(*args, **kwargs):
                if args != ():
                    print('called with args', args)
                if kwargs != {}:
                    print('called with kwargs', kwargs)
                if args == () and kwargs == {}:
                    print('called with no arguments')",
            "",
            "",
        )?
        .getattr("example")?
        .into();

        // call object with PyDict
        let kwargs = [(key1, val1)].into_py_dict_bound(py);
        fun.call_bound(py, (), Some(&kwargs))?;

        // pass arguments as Vec
        let kwargs = vec![(key1, val1), (key2, val2)];
        fun.call_bound(py, (), Some(&kwargs.into_py_dict_bound(py)))?;

        // pass arguments as HashMap
        let mut kwargs = HashMap::<&str, i32>::new();
        kwargs.insert(key1, 1);
        fun.call_bound(py, (), Some(&kwargs.into_py_dict_bound(py)))?;

        Ok(())
    })
}

Executing existing Python code

If you already have some existing Python code that you need to execute from Rust, the following FAQs can help you select the right PyO3 functionality for your situation:

Want to access Python APIs? Then use PyModule::import.

Pymodule::import can be used to get handle to a Python module from Rust. You can use this to import and use any Python module available in your environment.

use pyo3::prelude::*;

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let builtins = PyModule::import_bound(py, "builtins")?;
        let total: i32 = builtins
            .getattr("sum")?
            .call1((vec![1, 2, 3],))?
            .extract()?;
        assert_eq!(total, 6);
        Ok(())
    })
}

Want to run just an expression? Then use eval.

Python::eval is a method to execute a Python expression and return the evaluated value as a Bound<'py, PyAny> object.

use pyo3::prelude::*;

fn main() -> Result<(), ()> {
Python::with_gil(|py| {
    let result = py
        .eval_bound("[i * 10 for i in range(5)]", None, None)
        .map_err(|e| {
            e.print_and_set_sys_last_vars(py);
        })?;
    let res: Vec<i64> = result.extract().unwrap();
    assert_eq!(res, vec![0, 10, 20, 30, 40]);
    Ok(())
})
}

Want to run statements? Then use run.

Python::run is a method to execute one or more Python statements. This method returns nothing (like any Python statement), but you can get access to manipulated objects via the locals dict.

You can also use the py_run! macro, which is a shorthand for Python::run. Since py_run! panics on exceptions, we recommend you use this macro only for quickly testing your Python extensions.

use pyo3::prelude::*;
use pyo3::py_run;

fn main() {
#[pyclass]
struct UserData {
    id: u32,
    name: String,
}

#[pymethods]
impl UserData {
    fn as_tuple(&self) -> (u32, String) {
        (self.id, self.name.clone())
    }

    fn __repr__(&self) -> PyResult<String> {
        Ok(format!("User {}(id: {})", self.name, self.id))
    }
}

Python::with_gil(|py| {
    let userdata = UserData {
        id: 34,
        name: "Yu".to_string(),
    };
    let userdata = Py::new(py, userdata).unwrap();
    let userdata_as_tuple = (34, "Yu");
    py_run!(py, userdata userdata_as_tuple, r#"
assert repr(userdata) == "User Yu(id: 34)"
assert userdata.as_tuple() == userdata_as_tuple
    "#);
})
}

You have a Python file or code snippet? Then use PyModule::from_code.

PyModule::from_code can be used to generate a Python module which can then be used just as if it was imported with PyModule::import.

Warning: This will compile and execute code. Never pass untrusted code to this function!

use pyo3::{prelude::*, types::IntoPyDict};

fn main() -> PyResult<()> {
Python::with_gil(|py| {
    let activators = PyModule::from_code_bound(
        py,
        r#"
def relu(x):
    """see https://en.wikipedia.org/wiki/Rectifier_(neural_networks)"""
    return max(0.0, x)

def leaky_relu(x, slope=0.01):
    return x if x >= 0 else x * slope
    "#,
        "activators.py",
        "activators",
    )?;

    let relu_result: f64 = activators.getattr("relu")?.call1((-1.0,))?.extract()?;
    assert_eq!(relu_result, 0.0);

    let kwargs = [("slope", 0.2)].into_py_dict_bound(py);
    let lrelu_result: f64 = activators
        .getattr("leaky_relu")?
        .call((-1.0,), Some(&kwargs))?
        .extract()?;
    assert_eq!(lrelu_result, -0.2);
   Ok(())
})
}

Want to embed Python in Rust with additional modules?

Python maintains the sys.modules dict as a cache of all imported modules. An import in Python will first attempt to lookup the module from this dict, and if not present will use various strategies to attempt to locate and load the module.

The append_to_inittab macro can be used to add additional #[pymodule] modules to an embedded Python interpreter. The macro must be invoked before initializing Python.

As an example, the below adds the module foo to the embedded interpreter:

use pyo3::prelude::*;

#[pyfunction]
fn add_one(x: i64) -> i64 {
    x + 1
}

#[pymodule]
fn foo(foo_module: &Bound<'_, PyModule>) -> PyResult<()> {
    foo_module.add_function(wrap_pyfunction!(add_one, foo_module)?)?;
    Ok(())
}

fn main() -> PyResult<()> {
    pyo3::append_to_inittab!(foo);
    Python::with_gil(|py| Python::run_bound(py, "import foo; foo.add_one(6)", None, None))
}

If append_to_inittab cannot be used due to constraints in the program, an alternative is to create a module using PyModule::new and insert it manually into sys.modules:

use pyo3::prelude::*;
use pyo3::types::PyDict;

#[pyfunction]
pub fn add_one(x: i64) -> i64 {
    x + 1
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        // Create new module
        let foo_module = PyModule::new_bound(py, "foo")?;
        foo_module.add_function(wrap_pyfunction!(add_one, &foo_module)?)?;

        // Import and get sys.modules
        let sys = PyModule::import_bound(py, "sys")?;
        let py_modules: Bound<'_, PyDict> = sys.getattr("modules")?.downcast_into()?;

        // Insert foo into sys.modules
        py_modules.set_item("foo", foo_module)?;

        // Now we can import + run our python code
        Python::run_bound(py, "import foo; foo.add_one(6)", None, None)
    })
}

Include multiple Python files

You can include a file at compile time by using std::include_str macro.

Or you can load a file at runtime by using std::fs::read_to_string function.

Many Python files can be included and loaded as modules. If one file depends on another you must preserve correct order while declaring PyModule.

Example directory structure:

.
├── Cargo.lock
├── Cargo.toml
├── python_app
│   ├── app.py
│   └── utils
│       └── foo.py
└── src
    └── main.rs

python_app/app.py:

from utils.foo import bar


def run():
    return bar()

python_app/utils/foo.py:

def bar():
    return "baz"

The example below shows:

• how to include content of app.py and utils/foo.py into your rust binary
• how to call function run() (declared in app.py) that needs function imported from utils/foo.py

src/main.rs:

use pyo3::prelude::*;

fn main() -> PyResult<()> {
    let py_foo = include_str!(concat!(
        env!("CARGO_MANIFEST_DIR"),
        "/python_app/utils/foo.py"
    ));
    let py_app = include_str!(concat!(env!("CARGO_MANIFEST_DIR"), "/python_app/app.py"));
    let from_python = Python::with_gil(|py| -> PyResult<Py<PyAny>> {
        PyModule::from_code_bound(py, py_foo, "utils.foo", "utils.foo")?;
        let app: Py<PyAny> = PyModule::from_code_bound(py, py_app, "", "")?
            .getattr("run")?
            .into();
        app.call0(py)
    });

    println!("py: {}", from_python?);
    Ok(())
}

The example below shows:

• how to load content of app.py at runtime so that it sees its dependencies automatically
• how to call function run() (declared in app.py) that needs function imported from utils/foo.py

It is recommended to use absolute paths because then your binary can be run from anywhere as long as your app.py is in the expected directory (in this example that directory is /usr/share/python_app).

src/main.rs:

use pyo3::prelude::*;
use pyo3::types::PyList;
use std::fs;
use std::path::Path;

fn main() -> PyResult<()> {
    let path = Path::new("/usr/share/python_app");
    let py_app = fs::read_to_string(path.join("app.py"))?;
    let from_python = Python::with_gil(|py| -> PyResult<Py<PyAny>> {
        let syspath = py
            .import_bound("sys")?
            .getattr("path")?
            .downcast_into::<PyList>()?;
        syspath.insert(0, &path)?;
        let app: Py<PyAny> = PyModule::from_code_bound(py, &py_app, "", "")?
            .getattr("run")?
            .into();
        app.call0(py)
    });

    println!("py: {}", from_python?);
    Ok(())
}

Need to use a context manager from Rust?

Use context managers by directly invoking __enter__ and __exit__.

use pyo3::prelude::*;

fn main() {
    Python::with_gil(|py| {
        let custom_manager = PyModule::from_code_bound(
            py,
            r#"
class House(object):
    def __init__(self, address):
        self.address = address
    def __enter__(self):
        print(f"Welcome to {self.address}!")
    def __exit__(self, type, value, traceback):
        if type:
            print(f"Sorry you had {type} trouble at {self.address}")
        else:
            print(f"Thank you for visiting {self.address}, come again soon!")

        "#,
            "house.py",
            "house",
        )
        .unwrap();

        let house_class = custom_manager.getattr("House").unwrap();
        let house = house_class.call1(("123 Main Street",)).unwrap();

        house.call_method0("__enter__").unwrap();

        let result = py.eval_bound("undefined_variable + 1", None, None);

        // If the eval threw an exception we'll pass it through to the context manager.
        // Otherwise, __exit__  is called with empty arguments (Python "None").
        match result {
            Ok(_) => {
                let none = py.None();
                house
                    .call_method1("__exit__", (&none, &none, &none))
                    .unwrap();
            }
            Err(e) => {
                house
                    .call_method1(
                        "__exit__",
                        (
                            e.get_type_bound(py),
                            e.value_bound(py),
                            e.traceback_bound(py),
                        ),
                    )
                    .unwrap();
            }
        }
    })
}

Handling system signals/interrupts (Ctrl-C)

The best way to handle system signals when running Rust code is to periodically call Python::check_signals to handle any signals captured by Python's signal handler. See also the FAQ entry.

Alternatively, set Python's signal module to take the default action for a signal:

use pyo3::prelude::*;

fn main() -> PyResult<()> {
Python::with_gil(|py| -> PyResult<()> {
    let signal = py.import_bound("signal")?;
    // Set SIGINT to have the default action
    signal
        .getattr("signal")?
        .call1((signal.getattr("SIGINT")?, signal.getattr("SIG_DFL")?))?;
    Ok(())
})
}

Type conversions

In this portion of the guide we'll talk about the mapping of Python types to Rust types offered by PyO3, as well as the traits available to perform conversions between them.

Mapping of Rust types to Python types

When writing functions callable from Python (such as a #[pyfunction] or in a #[pymethods] block), the trait FromPyObject is required for function arguments, and IntoPy<PyObject> is required for function return values.

Consult the tables in the following section to find the Rust types provided by PyO3 which implement these traits.

Argument Types

When accepting a function argument, it is possible to either use Rust library types or PyO3's Python-native types. (See the next section for discussion on when to use each.)

The table below contains the Python type and the corresponding function argument types that will accept them:

It is also worth remembering the following special types:

For more detail on accepting #[pyclass] values as function arguments, see the section of this guide on Python Classes.

Using Rust library types vs Python-native types

Using Rust library types as function arguments will incur a conversion cost compared to using the Python-native types. Using the Python-native types is almost zero-cost (they just require a type check similar to the Python builtin function isinstance()).

However, once that conversion cost has been paid, the Rust standard library types offer a number of benefits:

• You can write functionality in native-speed Rust code (free of Python's runtime costs).
• You get better interoperability with the rest of the Rust ecosystem.
• You can use Python::allow_threads to release the Python GIL and let other Python threads make progress while your Rust code is executing.
• You also benefit from stricter type checking. For example you can specify Vec<i32>, which will only accept a Python list containing integers. The Python-native equivalent, &PyList, would accept a Python list containing Python objects of any type.

For most PyO3 usage the conversion cost is worth paying to get these benefits. As always, if you're not sure it's worth it in your case, benchmark it!

Returning Rust values to Python

When returning values from functions callable from Python, PyO3's smart pointers (Py<T>, Bound<'py, T>, and Borrowed<'a, 'py, T>) can be used with zero cost.

Because Bound<'py, T> and Borrowed<'a, 'py, T> have lifetime parameters, the Rust compiler may ask for lifetime annotations to be added to your function. See the section of the guide dedicated to this.

If your function is fallible, it should return PyResult<T> or Result<T, E> where E implements From<E> for PyErr. This will raise a Python exception if the Err variant is returned.

Finally, the following Rust types are also able to convert to Python as return values:

Conversion traits

PyO3 provides some handy traits to convert between Python types and Rust types.

.extract() and the FromPyObject trait

The easiest way to convert a Python object to a Rust value is using .extract(). It returns a PyResult with a type error if the conversion fails, so usually you will use something like

use pyo3::prelude::*;
use pyo3::types::PyList;
fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let list = PyList::new_bound(py, b"foo");
let v: Vec<i32> = list.extract()?;
        assert_eq!(&v, &[102, 111, 111]);
        Ok(())
    })
}

This method is available for many Python object types, and can produce a wide variety of Rust types, which you can check out in the implementor list of FromPyObject.

FromPyObject is also implemented for your own Rust types wrapped as Python objects (see the chapter about classes). There, in order to both be able to operate on mutable references and satisfy Rust's rules of non-aliasing mutable references, you have to extract the PyO3 reference wrappers PyRef and PyRefMut. They work like the reference wrappers of std::cell::RefCell and ensure (at runtime) that Rust borrows are allowed.

Deriving FromPyObject

FromPyObject can be automatically derived for many kinds of structs and enums if the member types themselves implement FromPyObject. This even includes members with a generic type T: FromPyObject. Derivation for empty enums, enum variants and structs is not supported.

Deriving FromPyObject for structs

The derivation generates code that will attempt to access the attribute my_string on the Python object, i.e. obj.getattr("my_string"), and call extract() on the attribute.

use pyo3::prelude::*;

#[derive(FromPyObject)]
struct RustyStruct {
    my_string: String,
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        let module = PyModule::from_code_bound(
            py,
            "class Foo:
            def __init__(self):
                self.my_string = 'test'",
            "",
            "",
        )?;

        let class = module.getattr("Foo")?;
        let instance = class.call0()?;
        let rustystruct: RustyStruct = instance.extract()?;
        assert_eq!(rustystruct.my_string, "test");
        Ok(())
    })
}

By setting the #[pyo3(item)] attribute on the field, PyO3 will attempt to extract the value by calling the get_item method on the Python object.

use pyo3::prelude::*;

#[derive(FromPyObject)]
struct RustyStruct {
    #[pyo3(item)]
    my_string: String,
}

use pyo3::types::PyDict;
fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        let dict = PyDict::new_bound(py);
        dict.set_item("my_string", "test")?;

        let rustystruct: RustyStruct = dict.extract()?;
        assert_eq!(rustystruct.my_string, "test");
        Ok(())
    })
}

The argument passed to getattr and get_item can also be configured:

use pyo3::prelude::*;

#[derive(FromPyObject)]
struct RustyStruct {
    #[pyo3(item("key"))]
    string_in_mapping: String,
    #[pyo3(attribute("name"))]
    string_attr: String,
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        let module = PyModule::from_code_bound(
            py,
            "class Foo(dict):
            def __init__(self):
                self.name = 'test'
                self['key'] = 'test2'",
            "",
            "",
        )?;

        let class = module.getattr("Foo")?;
        let instance = class.call0()?;
        let rustystruct: RustyStruct = instance.extract()?;
		assert_eq!(rustystruct.string_attr, "test");
        assert_eq!(rustystruct.string_in_mapping, "test2");

        Ok(())
    })
}

This tries to extract string_attr from the attribute name and string_in_mapping from a mapping with the key "key". The arguments for attribute are restricted to non-empty string literals while item can take any valid literal that implements ToBorrowedObject.

You can use #[pyo3(from_item_all)] on a struct to extract every field with get_item method. In this case, you can't use #[pyo3(attribute)] or barely use #[pyo3(item)] on any field. However, using #[pyo3(item("key"))] to specify the key for a field is still allowed.

use pyo3::prelude::*;

#[derive(FromPyObject)]
#[pyo3(from_item_all)]
struct RustyStruct {
    foo: String,
    bar: String,
    #[pyo3(item("foobar"))]
    baz: String,
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        let py_dict = py.eval_bound("{'foo': 'foo', 'bar': 'bar', 'foobar': 'foobar'}", None, None)?;
        let rustystruct: RustyStruct = py_dict.extract()?;
		  assert_eq!(rustystruct.foo, "foo");
        assert_eq!(rustystruct.bar, "bar");
        assert_eq!(rustystruct.baz, "foobar");

        Ok(())
    })
}

Deriving FromPyObject for tuple structs

Tuple structs are also supported but do not allow customizing the extraction. The input is always assumed to be a Python tuple with the same length as the Rust type, the nth field is extracted from the nth item in the Python tuple.

use pyo3::prelude::*;

#[derive(FromPyObject)]
struct RustyTuple(String, String);

use pyo3::types::PyTuple;
fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        let tuple = PyTuple::new_bound(py, vec!["test", "test2"]);

        let rustytuple: RustyTuple = tuple.extract()?;
        assert_eq!(rustytuple.0, "test");
        assert_eq!(rustytuple.1, "test2");

        Ok(())
    })
}

Tuple structs with a single field are treated as wrapper types which are described in the following section. To override this behaviour and ensure that the input is in fact a tuple, specify the struct as

use pyo3::prelude::*;

#[derive(FromPyObject)]
struct RustyTuple((String,));

use pyo3::types::PyTuple;
fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        let tuple = PyTuple::new_bound(py, vec!["test"]);

        let rustytuple: RustyTuple = tuple.extract()?;
        assert_eq!((rustytuple.0).0, "test");

        Ok(())
    })
}

Deriving FromPyObject for wrapper types

The pyo3(transparent) attribute can be used on structs with exactly one field. This results in extracting directly from the input object, i.e. obj.extract(), rather than trying to access an item or attribute. This behaviour is enabled per default for newtype structs and tuple-variants with a single field.

use pyo3::prelude::*;

#[derive(FromPyObject)]
struct RustyTransparentTupleStruct(String);

#[derive(FromPyObject)]
#[pyo3(transparent)]
struct RustyTransparentStruct {
    inner: String,
}

use pyo3::types::PyString;
fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        let s = PyString::new_bound(py, "test");

        let tup: RustyTransparentTupleStruct = s.extract()?;
        assert_eq!(tup.0, "test");

        let stru: RustyTransparentStruct = s.extract()?;
        assert_eq!(stru.inner, "test");

        Ok(())
    })
}

Deriving FromPyObject for enums

The FromPyObject derivation for enums generates code that tries to extract the variants in the order of the fields. As soon as a variant can be extracted successfully, that variant is returned. This makes it possible to extract Python union types like str | int.

The same customizations and restrictions described for struct derivations apply to enum variants, i.e. a tuple variant assumes that the input is a Python tuple, and a struct variant defaults to extracting fields as attributes but can be configured in the same manner. The transparent attribute can be applied to single-field-variants.

use pyo3::prelude::*;

#[derive(FromPyObject)]
#[derive(Debug)]
enum RustyEnum<'a> {
    Int(usize),                    // input is a positive int
    String(String),                // input is a string
    IntTuple(usize, usize),        // input is a 2-tuple with positive ints
    StringIntTuple(String, usize), // input is a 2-tuple with String and int
    Coordinates3d {
        // needs to be in front of 2d
        x: usize,
        y: usize,
        z: usize,
    },
    Coordinates2d {
        // only gets checked if the input did not have `z`
        #[pyo3(attribute("x"))]
        a: usize,
        #[pyo3(attribute("y"))]
        b: usize,
    },
    #[pyo3(transparent)]
    CatchAll(&'a PyAny), // This extraction never fails
}

use pyo3::types::{PyBytes, PyString};
fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        {
            let thing = 42_u8.to_object(py);
            let rust_thing: RustyEnum<'_> = thing.extract(py)?;

            assert_eq!(
                42,
                match rust_thing {
                    RustyEnum::Int(i) => i,
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }
        {
            let thing = PyString::new_bound(py, "text");
            let rust_thing: RustyEnum<'_> = thing.extract()?;

            assert_eq!(
                "text",
                match rust_thing {
                    RustyEnum::String(i) => i,
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }
        {
            let thing = (32_u8, 73_u8).to_object(py);
            let rust_thing: RustyEnum<'_> = thing.extract(py)?;

            assert_eq!(
                (32, 73),
                match rust_thing {
                    RustyEnum::IntTuple(i, j) => (i, j),
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }
        {
            let thing = ("foo", 73_u8).to_object(py);
            let rust_thing: RustyEnum<'_> = thing.extract(py)?;

            assert_eq!(
                (String::from("foo"), 73),
                match rust_thing {
                    RustyEnum::StringIntTuple(i, j) => (i, j),
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }
        {
            let module = PyModule::from_code_bound(
                py,
                "class Foo(dict):
            def __init__(self):
                self.x = 0
                self.y = 1
                self.z = 2",
                "",
                "",
            )?;

            let class = module.getattr("Foo")?;
            let instance = class.call0()?;
            let rust_thing: RustyEnum<'_> = instance.extract()?;

            assert_eq!(
                (0, 1, 2),
                match rust_thing {
                    RustyEnum::Coordinates3d { x, y, z } => (x, y, z),
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }

        {
            let module = PyModule::from_code_bound(
                py,
                "class Foo(dict):
            def __init__(self):
                self.x = 3
                self.y = 4",
                "",
                "",
            )?;

            let class = module.getattr("Foo")?;
            let instance = class.call0()?;
            let rust_thing: RustyEnum<'_> = instance.extract()?;

            assert_eq!(
                (3, 4),
                match rust_thing {
                    RustyEnum::Coordinates2d { a, b } => (a, b),
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }

        {
            let thing = PyBytes::new_bound(py, b"text");
            let rust_thing: RustyEnum<'_> = thing.extract()?;

            assert_eq!(
                b"text",
                match rust_thing {
                    RustyEnum::CatchAll(i) => i.downcast::<PyBytes>()?.as_bytes(),
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }
        Ok(())
    })
}

If none of the enum variants match, a PyTypeError containing the names of the tested variants is returned. The names reported in the error message can be customized through the #[pyo3(annotation = "name")] attribute, e.g. to use conventional Python type names:

use pyo3::prelude::*;

#[derive(FromPyObject)]
#[derive(Debug)]
enum RustyEnum {
    #[pyo3(transparent, annotation = "str")]
    String(String),
    #[pyo3(transparent, annotation = "int")]
    Int(isize),
}

fn main() -> PyResult<()> {
    Python::with_gil(|py| -> PyResult<()> {
        {
            let thing = 42_u8.to_object(py);
            let rust_thing: RustyEnum = thing.extract(py)?;

            assert_eq!(
                42,
                match rust_thing {
                    RustyEnum::Int(i) => i,
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }

        {
            let thing = "foo".to_object(py);
            let rust_thing: RustyEnum = thing.extract(py)?;

            assert_eq!(
                "foo",
                match rust_thing {
                    RustyEnum::String(i) => i,
                    other => unreachable!("Error extracting: {:?}", other),
                }
            );
        }

        {
            let thing = b"foo".to_object(py);
            let error = thing.extract::<RustyEnum>(py).unwrap_err();
            assert!(error.is_instance_of::<pyo3::exceptions::PyTypeError>(py));
        }

        Ok(())
    })
}

If the input is neither a string nor an integer, the error message will be: "'<INPUT_TYPE>' cannot be converted to 'str | int'".

#[derive(FromPyObject)] Container Attributes

• pyo3(transparent)
  • extract the field directly from the object as obj.extract() instead of get_item() or getattr()
  • Newtype structs and tuple-variants are treated as transparent per default.
  • only supported for single-field structs and enum variants
• pyo3(annotation = "name")
  • changes the name of the failed variant in the generated error message in case of failure.
  • e.g. pyo3("int") reports the variant's type as int.
  • only supported for enum variants

#[derive(FromPyObject)] Field Attributes

• pyo3(attribute), pyo3(attribute("name"))
  • retrieve the field from an attribute, possibly with a custom name specified as an argument
  • argument must be a string-literal.
• pyo3(item), pyo3(item("key"))
  • retrieve the field from a mapping, possibly with the custom key specified as an argument.
  • can be any literal that implements ToBorrowedObject
• pyo3(from_py_with = "...")
  • apply a custom function to convert the field from Python the desired Rust type.
  • the argument must be the name of the function as a string.
  • the function signature must be fn(&Bound<PyAny>) -> PyResult<T> where T is the Rust type of the argument.

IntoPy<T>

This trait defines the to-python conversion for a Rust type. It is usually implemented as IntoPy<PyObject>, which is the trait needed for returning a value from #[pyfunction] and #[pymethods].

All types in PyO3 implement this trait, as does a #[pyclass] which doesn't use extends.

Occasionally you may choose to implement this for custom types which are mapped to Python types without having a unique python type.

use pyo3::prelude::*;
#[allow(dead_code)]
struct MyPyObjectWrapper(PyObject);

impl IntoPy<PyObject> for MyPyObjectWrapper {
    fn into_py(self, py: Python<'_>) -> PyObject {
        self.0
    }
}

The ToPyObject trait

ToPyObject is a conversion trait that allows various objects to be converted into PyObject. IntoPy<PyObject> serves the same purpose, except that it consumes self.

Using async and await

This feature is still in active development. See the related issue.

#[pyfunction] and #[pymethods] attributes also support async fn.

#![allow(dead_code)]
#[cfg(feature = "experimental-async")] {
use std::{thread, time::Duration};
use futures::channel::oneshot;
use pyo3::prelude::*;

#[pyfunction]
async fn sleep(seconds: f64, result: Option<PyObject>) -> Option<PyObject> {
    let (tx, rx) = oneshot::channel();
    thread::spawn(move || {
        thread::sleep(Duration::from_secs_f64(seconds));
        tx.send(()).unwrap();
    });
    rx.await.unwrap();
    result
}
}

Python awaitables instantiated with this method can only be awaited in asyncio context. Other Python async runtime may be supported in the future.

Send + 'static constraint

Resulting future of an async fn decorated by #[pyfunction] must be Send + 'static to be embedded in a Python object.

As a consequence, async fn parameters and return types must also be Send + 'static, so it is not possible to have a signature like async fn does_not_compile<'py>(arg: Bound<'py, PyAny>) -> Bound<'py, PyAny>.

However, there is an exception for method receivers, so async methods can accept &self/&mut self. Note that this means that the class instance is borrowed for as long as the returned future is not completed, even across yield points and while waiting for I/O operations to complete. Hence, other methods cannot obtain exclusive borrows while the future is still being polled. This is the same as how async methods in Rust generally work but it is more problematic for Rust code interfacing with Python code due to pervasive shared mutability. This strongly suggests to prefer shared borrows &self over exclusive ones &mut self to avoid racy borrow check failures at runtime.

Implicit GIL holding

Even if it is not possible to pass a py: Python<'py> parameter to async fn, the GIL is still held during the execution of the future – it's also the case for regular fn without Python<'py>/Bound<'py, PyAny> parameter, yet the GIL is held.

It is still possible to get a Python marker using Python::with_gil; because with_gil is reentrant and optimized, the cost will be negligible.

Release the GIL across .await

There is currently no simple way to release the GIL when awaiting a future, but solutions are currently in development.

Here is the advised workaround for now:

use std::{
    future::Future,
    pin::{Pin, pin},
    task::{Context, Poll},
};
use pyo3::prelude::*;

struct AllowThreads<F>(F);

impl<F> Future for AllowThreads<F>
where
    F: Future + Unpin + Send,
    F::Output: Send,
{
    type Output = F::Output;

    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        let waker = cx.waker();
        Python::with_gil(|gil| {
            gil.allow_threads(|| pin!(&mut self.0).poll(&mut Context::from_waker(waker)))
        })
    }
}

Cancellation

Cancellation on the Python side can be caught using CancelHandle type, by annotating a function parameter with #[pyo3(cancel_handle)].

#![allow(dead_code)]
#[cfg(feature = "experimental-async")] {
use futures::FutureExt;
use pyo3::prelude::*;
use pyo3::coroutine::CancelHandle;

#[pyfunction]
async fn cancellable(#[pyo3(cancel_handle)] mut cancel: CancelHandle) {
    futures::select! {
        /* _ = ... => println!("done"), */
        _ = cancel.cancelled().fuse() => println!("cancelled"),
    }
}
}

The Coroutine type

To make a Rust future awaitable in Python, PyO3 defines a Coroutine type, which implements the Python coroutine protocol.

Each coroutine.send call is translated to a Future::poll call. If a CancelHandle parameter is declared, the exception passed to coroutine.throw call is stored in it and can be retrieved with CancelHandle::cancelled; otherwise, it cancels the Rust future, and the exception is reraised;

The type does not yet have a public constructor until the design is finalized.

Parallelism

CPython has the infamous Global Interpreter Lock, which prevents several threads from executing Python bytecode in parallel. This makes threading in Python a bad fit for CPU-bound tasks and often forces developers to accept the overhead of multiprocessing.

In PyO3 parallelism can be easily achieved in Rust-only code. Let's take a look at our word-count example, where we have a search function that utilizes the rayon crate to count words in parallel.

#![allow(dead_code)]
use pyo3::prelude::*;

// These traits let us use `par_lines` and `map`.
use rayon::str::ParallelString;
use rayon::iter::ParallelIterator;

/// Count the occurrences of needle in line, case insensitive
fn count_line(line: &str, needle: &str) -> usize {
    let mut total = 0;
    for word in line.split(' ') {
        if word == needle {
            total += 1;
        }
    }
    total
}

#[pyfunction]
fn search(contents: &str, needle: &str) -> usize {
    contents
        .par_lines()
        .map(|line| count_line(line, needle))
        .sum()
}

But let's assume you have a long running Rust function which you would like to execute several times in parallel. For the sake of example let's take a sequential version of the word count:

#![allow(dead_code)]
fn count_line(line: &str, needle: &str) -> usize {
    let mut total = 0;
    for word in line.split(' ') {
        if word == needle {
            total += 1;
        }
    }
    total
}

fn search_sequential(contents: &str, needle: &str) -> usize {
    contents.lines().map(|line| count_line(line, needle)).sum()
}

To enable parallel execution of this function, the Python::allow_threads method can be used to temporarily release the GIL, thus allowing other Python threads to run. We then have a function exposed to the Python runtime which calls search_sequential inside a closure passed to Python::allow_threads to enable true parallelism:

#![allow(dead_code)]
use pyo3::prelude::*;

fn count_line(line: &str, needle: &str) -> usize {
    let mut total = 0;
    for word in line.split(' ') {
        if word == needle {
            total += 1;
        }
    }
    total
}

fn search_sequential(contents: &str, needle: &str) -> usize {
   contents.lines().map(|line| count_line(line, needle)).sum()
}
#[pyfunction]
fn search_sequential_allow_threads(py: Python<'_>, contents: &str, needle: &str) -> usize {
    py.allow_threads(|| search_sequential(contents, needle))
}

Now Python threads can use more than one CPU core, resolving the limitation which usually makes multi-threading in Python only good for IO-bound tasks:

from concurrent.futures import ThreadPoolExecutor
from word_count import search_sequential_allow_threads

executor = ThreadPoolExecutor(max_workers=2)

future_1 = executor.submit(
    word_count.search_sequential_allow_threads, contents, needle
)
future_2 = executor.submit(
    word_count.search_sequential_allow_threads, contents, needle
)
result_1 = future_1.result()
result_2 = future_2.result()

Benchmark

Let's benchmark the word-count example to verify that we really did unlock parallelism with PyO3.

We are using pytest-benchmark to benchmark four word count functions:

1. Pure Python version
2. Rust parallel version
3. Rust sequential version
4. Rust sequential version executed twice with two Python threads

The benchmark script can be found here, and we can run nox in the word-count folder to benchmark these functions.

While the results of the benchmark of course depend on your machine, the relative results should be similar to this (mid 2020):

-------------------------------------------------------------------------------------------------- benchmark: 4 tests -------------------------------------------------------------------------------------------------
Name (time in ms)                                          Min                Max               Mean            StdDev             Median               IQR            Outliers       OPS            Rounds  Iterations
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_word_count_rust_parallel                           1.7315 (1.0)       4.6495 (1.0)       1.9972 (1.0)      0.4299 (1.0)       1.8142 (1.0)      0.2049 (1.0)         40;46  500.6943 (1.0)         375           1
test_word_count_rust_sequential                         7.3348 (4.24)     10.3556 (2.23)      8.0035 (4.01)     0.7785 (1.81)      7.5597 (4.17)     0.8641 (4.22)         26;5  124.9457 (0.25)        121           1
test_word_count_rust_sequential_twice_with_threads      7.9839 (4.61)     10.3065 (2.22)      8.4511 (4.23)     0.4709 (1.10)      8.2457 (4.55)     0.3927 (1.92)        17;17  118.3274 (0.24)        114           1
test_word_count_python_sequential                      27.3985 (15.82)    45.4527 (9.78)     28.9604 (14.50)    4.1449 (9.64)     27.5781 (15.20)    0.4638 (2.26)          3;5   34.5299 (0.07)         35           1
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

You can see that the Python threaded version is not much slower than the Rust sequential version, which means compared to an execution on a single CPU core the speed has doubled.

Debugging

Macros

PyO3's attributes (#[pyclass], #[pymodule], etc.) are procedural macros, which means that they rewrite the source of the annotated item. You can view the generated source with the following command, which also expands a few other things:

cargo rustc --profile=check -- -Z unstable-options --pretty=expanded > expanded.rs; rustfmt expanded.rs

(You might need to install rustfmt if you don't already have it.)

You can also debug classic !-macros by adding -Z trace-macros:

cargo rustc --profile=check -- -Z unstable-options --pretty=expanded -Z trace-macros > expanded.rs; rustfmt expanded.rs

Note that those commands require using the nightly build of rust and may occasionally have bugs. See cargo expand for a more elaborate and stable version of those commands.

Running with Valgrind

Valgrind is a tool to detect memory management bugs such as memory leaks.

You first need to install a debug build of Python, otherwise Valgrind won't produce usable results. In Ubuntu there's e.g. a python3-dbg package.

Activate an environment with the debug interpreter and recompile. If you're on Linux, use ldd with the name of your binary and check that you're linking e.g. libpython3.7d.so.1.0 instead of libpython3.7.so.1.0.

Download the suppressions file for CPython.

Run Valgrind with valgrind --suppressions=valgrind-python.supp ./my-command --with-options

Getting a stacktrace

The best start to investigate a crash such as an segmentation fault is a backtrace. You can set RUST_BACKTRACE=1 as an environment variable to get the stack trace on a panic!. Alternatively you can use a debugger such as gdb to explore the issue. Rust provides a wrapper, rust-gdb, which has pretty-printers for inspecting Rust variables. Since PyO3 uses cdylib for Python shared objects, it does not receive the pretty-print debug hooks in rust-gdb (rust-lang/rust#96365). The mentioned issue contains a workaround for enabling pretty-printers in this case.

• Link against a debug build of python as described in the previous chapter
• Run rust-gdb <my-binary>
• Set a breakpoint (b) on rust_panic if you are investigating a panic!
• Enter r to run
• After the crash occurred, enter bt or bt full to print the stacktrace

Often it is helpful to run a small piece of Python code to exercise a section of Rust.

rust-gdb --args python -c "import my_package; my_package.sum_to_string(1, 2)"

Features reference

PyO3 provides a number of Cargo features to customize functionality. This chapter of the guide provides detail on each of them.

By default, only the macros feature is enabled.

Features for extension module authors

extension-module

This feature is required when building a Python extension module using PyO3.

It tells PyO3's build script to skip linking against libpython.so on Unix platforms, where this must not be done.

See the building and distribution section for further detail.

abi3

This feature is used when building Python extension modules to create wheels which are compatible with multiple Python versions.

It restricts PyO3's API to a subset of the full Python API which is guaranteed by PEP 384 to be forwards-compatible with future Python versions.

See the building and distribution section for further detail.

The abi3-pyXY features

(abi3-py37, abi3-py38, abi3-py39, abi3-py310 and abi3-py311)

These features are extensions of the abi3 feature to specify the exact minimum Python version which the multiple-version-wheel will support.

See the building and distribution section for further detail.

generate-import-lib

This experimental feature is used to generate import libraries for Python DLL for MinGW-w64 and MSVC (cross-)compile targets.

Enabling it allows to (cross-)compile extension modules to any Windows targets without having to install the Windows Python distribution files for the target.

See the building and distribution section for further detail.

Features for embedding Python in Rust

auto-initialize

This feature changes Python::with_gil to automatically initialize a Python interpreter (by calling prepare_freethreaded_python) if needed.

If you do not enable this feature, you should call pyo3::prepare_freethreaded_python() before attempting to call any other Python APIs.

Advanced Features

experimental-async

This feature adds support for async fn in #[pyfunction] and #[pymethods].

The feature has some unfinished refinements and performance improvements. To help finish this off, see issue #1632 and its associated draft PRs.

experimental-declarative-modules

This feature allows to declare Python modules using #[pymodule] mod my_module { ... } syntax.

The feature has some unfinished refinements and edge cases. To help finish this off, see issue #3900.

experimental-inspect

This feature adds the pyo3::inspect module, as well as IntoPy::type_output and FromPyObject::type_input APIs to produce Python type "annotations" for Rust types.

This is a first step towards adding first-class support for generating type annotations automatically in PyO3, however work is needed to finish this off. All feedback and offers of help welcome on issue #2454.

gil-refs

This feature is a backwards-compatibility feature to allow continued use of the "GIL Refs" APIs deprecated in PyO3 0.21. These APIs have performance drawbacks and soundness edge cases which the newer Bound<T> smart pointer and accompanying APIs resolve.

This feature and the APIs it enables is expected to be removed in a future PyO3 version.

macros

This feature enables a dependency on the pyo3-macros crate, which provides the procedural macros portion of PyO3's API:

• #[pymodule]
• #[pyfunction]
• #[pyclass]
• #[pymethods]
• #[derive(FromPyObject)]

It also provides the py_run! macro.

These macros require a number of dependencies which may not be needed by users who just need PyO3 for Python FFI. Disabling this feature enables faster builds for those users, as these dependencies will not be built if this feature is disabled.

This feature is enabled by default. To disable it, set default-features = false for the pyo3 entry in your Cargo.toml.

multiple-pymethods

This feature enables a dependency on inventory, which enables each #[pyclass] to have more than one #[pymethods] block. This feature also requires a minimum Rust version of 1.62 due to limitations in the inventory crate.

Most users should only need a single #[pymethods] per #[pyclass]. In addition, not all platforms (e.g. Wasm) are supported by inventory. For this reason this feature is not enabled by default, meaning fewer dependencies and faster compilation for the majority of users.

See the #[pyclass] implementation details for more information.

nightly

The nightly feature needs the nightly Rust compiler. This allows PyO3 to use the auto_traits and negative_impls features to fix the Python::allow_threads function.

resolve-config

The resolve-config feature of the pyo3-build-config crate controls whether that crate's build script automatically resolves a Python interpreter / build configuration. This feature is primarily useful when building PyO3 itself. By default this feature is not enabled, meaning you can freely use pyo3-build-config as a standalone library to read or write PyO3 build configuration files or resolve metadata about a Python interpreter.

Optional Dependencies

These features enable conversions between Python types and types from other Rust crates, enabling easy access to the rest of the Rust ecosystem.

anyhow

Adds a dependency on anyhow. Enables a conversion from anyhow’s Error type to PyErr, for easy error handling.

chrono

Adds a dependency on chrono. Enables a conversion from chrono's types to python:

• TimeDelta -> PyDelta
• FixedOffset -> PyDelta
• Utc -> PyTzInfo
• NaiveDate -> PyDate
• NaiveTime -> PyTime
• DateTime -> PyDateTime

chrono-tz

Adds a dependency on chrono-tz. Enables conversion from and to Tz. It requires at least Python 3.9.

either

Adds a dependency on either. Enables a conversions into either’s Either type.

eyre

Adds a dependency on eyre. Enables a conversion from eyre’s Report type to PyErr, for easy error handling.

hashbrown

Adds a dependency on hashbrown and enables conversions into its HashMap and HashSet types.

indexmap

Adds a dependency on indexmap and enables conversions into its IndexMap type.

num-bigint

Adds a dependency on num-bigint and enables conversions into its BigInt and BigUint types.

num-complex

Adds a dependency on num-complex and enables conversions into its Complex type.

rust_decimal

Adds a dependency on rust_decimal and enables conversions into its Decimal type.

serde

Enables (de)serialization of Py<T> objects via serde. This allows to use #[derive(Serialize, Deserialize) on structs that hold references to #[pyclass] instances

#[cfg(feature = "serde")]
#[allow(dead_code)]
mod serde_only {
use pyo3::prelude::*;
use serde::{Deserialize, Serialize};

#[pyclass]
#[derive(Serialize, Deserialize)]
struct Permission {
    name: String,
}

#[pyclass]
#[derive(Serialize, Deserialize)]
struct User {
    username: String,
    permissions: Vec<Py<Permission>>,
}
}

smallvec

Adds a dependency on smallvec and enables conversions into its SmallVec type.

Memory management

Rust and Python have very different notions of memory management. Rust has a strict memory model with concepts of ownership, borrowing, and lifetimes, where memory is freed at predictable points in program execution. Python has a looser memory model in which variables are reference-counted with shared, mutable state by default. A global interpreter lock (GIL) is needed to prevent race conditions, and a garbage collector is needed to break reference cycles. Memory in Python is freed eventually by the garbage collector, but not usually in a predictable way.

PyO3 bridges the Rust and Python memory models with two different strategies for accessing memory allocated on Python's heap from inside Rust. These are GIL Refs such as &'py PyAny, and GIL-independent Py<Any> smart pointers.

GIL-bound memory

PyO3's GIL Refs such as &'py PyAny make PyO3 more ergonomic to use by ensuring that their lifetime can never be longer than the duration the Python GIL is held. This means that most of PyO3's API can assume the GIL is held. (If PyO3 could not assume this, every PyO3 API would need to take a Python GIL token to prove that the GIL is held.) This allows us to write very simple and easy-to-understand programs like this:

use pyo3::prelude::*;
use pyo3::types::PyString;
fn main() -> PyResult<()> {
Python::with_gil(|py| -> PyResult<()> {
    #[allow(deprecated)] // py.eval() is part of the GIL Refs API
    let hello = py
        .eval("\"Hello World!\"", None, None)?
        .downcast::<PyString>()?;
    println!("Python says: {}", hello);
    Ok(())
})?;
Ok(())
}

Internally, calling Python::with_gil() creates a GILPool which owns the memory pointed to by the reference. In the example above, the lifetime of the reference hello is bound to the GILPool. When the with_gil() closure ends the GILPool is also dropped and the Python reference counts of the variables it owns are decreased, releasing them to the Python garbage collector. Most of the time we don't have to think about this, but consider the following:

use pyo3::prelude::*;
use pyo3::types::PyString;
fn main() -> PyResult<()> {
Python::with_gil(|py| -> PyResult<()> {
    for _ in 0..10 {
        #[allow(deprecated)] // py.eval() is part of the GIL Refs API
        let hello = py
            .eval("\"Hello World!\"", None, None)?
            .downcast::<PyString>()?;
        println!("Python says: {}", hello);
    }
    // There are 10 copies of `hello` on Python's heap here.
    Ok(())
})?;
Ok(())
}

We might assume that the hello variable's memory is freed at the end of each loop iteration, but in fact we create 10 copies of hello on Python's heap. This may seem surprising at first, but it is completely consistent with Rust's memory model. The hello variable is dropped at the end of each loop, but it is only a reference to the memory owned by the GILPool, and its lifetime is bound to the GILPool, not the for loop. The GILPool isn't dropped until the end of the with_gil() closure, at which point the 10 copies of hello are finally released to the Python garbage collector.

In general we don't want unbounded memory growth during loops! One workaround is to acquire and release the GIL with each iteration of the loop.

use pyo3::prelude::*;
use pyo3::types::PyString;
fn main() -> PyResult<()> {
for _ in 0..10 {
    Python::with_gil(|py| -> PyResult<()> {
        #[allow(deprecated)] // py.eval() is part of the GIL Refs API
        let hello = py
            .eval("\"Hello World!\"", None, None)?
            .downcast::<PyString>()?;
        println!("Python says: {}", hello);
        Ok(())
    })?; // only one copy of `hello` at a time
}
Ok(())
}

It might not be practical or performant to acquire and release the GIL so many times. Another workaround is to work with the GILPool object directly, but this is unsafe.

use pyo3::prelude::*;
use pyo3::types::PyString;
fn main() -> PyResult<()> {
Python::with_gil(|py| -> PyResult<()> {
    for _ in 0..10 {
        #[allow(deprecated)] // `new_pool` is not needed in code not using the GIL Refs API
        let pool = unsafe { py.new_pool() };
        let py = pool.python();
        #[allow(deprecated)] // py.eval() is part of the GIL Refs API
        let hello = py
            .eval("\"Hello World!\"", None, None)?
            .downcast::<PyString>()?;
        println!("Python says: {}", hello);
    }
    Ok(())
})?;
Ok(())
}

The unsafe method Python::new_pool allows you to create a nested GILPool from which you can retrieve a new py: Python GIL token. Variables created with this new GIL token are bound to the nested GILPool and will be released when the nested GILPool is dropped. Here, the nested GILPool is dropped at the end of each loop iteration, before the with_gil() closure ends.

When doing this, you must be very careful to ensure that once the GILPool is dropped you do not retain access to any owned references created after the GILPool was created. Read the documentation for Python::new_pool() for more information on safety.

This memory management can also be applicable when writing extension modules. #[pyfunction] and #[pymethods] will create a GILPool which lasts the entire function call, releasing objects when the function returns. Most functions only create a few objects, meaning this doesn't have a significant impact. Occasionally functions with long complex loops may need to use Python::new_pool as shown above.

GIL-independent memory

Sometimes we need a reference to memory on Python's heap that can outlive the GIL. Python's Py<PyAny> is analogous to Arc<T>, but for variables whose memory is allocated on Python's heap. Cloning a Py<PyAny> increases its internal reference count just like cloning Arc<T>. The smart pointer can outlive the "GIL is held" period in which it was created. It isn't magic, though. We need to reacquire the GIL to access the memory pointed to by the Py<PyAny>.

What happens to the memory when the last Py<PyAny> is dropped and its reference count reaches zero? It depends whether or not we are holding the GIL.

use pyo3::prelude::*;
use pyo3::types::PyString;
fn main() -> PyResult<()> {
Python::with_gil(|py| -> PyResult<()> {
    #[allow(deprecated)] // py.eval() is part of the GIL Refs API
    let hello: Py<PyString> = py.eval("\"Hello World!\"", None, None)?.extract()?;
    #[allow(deprecated)] // as_ref is part of the GIL Refs API
    {
        println!("Python says: {}", hello.as_ref(py));
    }
    Ok(())
})?;
Ok(())
}

At the end of the Python::with_gil() closure hello is dropped, and then the GIL is dropped. Since hello is dropped while the GIL is still held by the current thread, its memory is released to the Python garbage collector immediately.

This example wasn't very interesting. We could have just used a GIL-bound &PyString reference. What happens when the last Py<Any> is dropped while we are not holding the GIL?

use pyo3::prelude::*;
use pyo3::types::PyString;
fn main() -> PyResult<()> {
let hello: Py<PyString> = Python::with_gil(|py| {
    #[allow(deprecated)] // py.eval() is part of the GIL Refs API
    py.eval("\"Hello World!\"", None, None)?.extract()
})?;
// Do some stuff...
// Now sometime later in the program we want to access `hello`.
Python::with_gil(|py| {
    #[allow(deprecated)]  // as_ref is part of the deprecated "GIL Refs" API.
    let hello = hello.as_ref(py);
    println!("Python says: {}", hello);
});
// Now we're done with `hello`.
drop(hello); // Memory *not* released here.
// Sometime later we need the GIL again for something...
Python::with_gil(|py|
    // Memory for `hello` is released here.
()
);
Ok(())
}

When hello is dropped nothing happens to the pointed-to memory on Python's heap because nothing can happen if we're not holding the GIL. Fortunately, the memory isn't leaked. PyO3 keeps track of the memory internally and will release it the next time we acquire the GIL.

We can avoid the delay in releasing memory if we are careful to drop the Py<Any> while the GIL is held.

use pyo3::prelude::*;
use pyo3::types::PyString;
fn main() -> PyResult<()> {
#[allow(deprecated)] // py.eval() is part of the GIL Refs API
let hello: Py<PyString> =
    Python::with_gil(|py| py.eval("\"Hello World!\"", None, None)?.extract())?;
// Do some stuff...
// Now sometime later in the program:
Python::with_gil(|py| {
    #[allow(deprecated)] // as_ref is part of the GIL Refs API
    {
        println!("Python says: {}", hello.as_ref(py));
    }
    drop(hello); // Memory released here.
});
Ok(())
}

We could also have used Py::into_ref(), which consumes self, instead of Py::as_ref(). But note that in addition to being slower than as_ref(), into_ref() binds the memory to the lifetime of the GILPool, which means that rather than being released immediately, the memory will not be released until the GIL is dropped.

use pyo3::prelude::*;
use pyo3::types::PyString;
fn main() -> PyResult<()> {
#[allow(deprecated)] // py.eval() is part of the GIL Refs API
let hello: Py<PyString> =
    Python::with_gil(|py| py.eval("\"Hello World!\"", None, None)?.extract())?;
// Do some stuff...
// Now sometime later in the program:
Python::with_gil(|py| {
    #[allow(deprecated)] // into_ref is part of the GIL Refs API
    {
        println!("Python says: {}", hello.into_ref(py));
    }
    // Memory not released yet.
    // Do more stuff...
    // Memory released here at end of `with_gil()` closure.
});
Ok(())
}

Performance

To achieve the best possible performance, it is useful to be aware of several tricks and sharp edges concerning PyO3's API.

extract versus downcast

Pythonic API implemented using PyO3 are often polymorphic, i.e. they will accept &Bound<'_, PyAny> and try to turn this into multiple more concrete types to which the requested operation is applied. This often leads to chains of calls to extract, e.g.

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::{exceptions::PyTypeError, types::PyList};

fn frobnicate_list<'py>(list: &Bound<'_, PyList>) -> PyResult<Bound<'py, PyAny>> {
    todo!()
}

fn frobnicate_vec<'py>(vec: Vec<Bound<'py, PyAny>>) -> PyResult<Bound<'py, PyAny>> {
    todo!()
}

#[pyfunction]
fn frobnicate<'py>(value: &Bound<'py, PyAny>) -> PyResult<Bound<'py, PyAny>> {
    if let Ok(list) = value.extract::<Bound<'_, PyList>>() {
        frobnicate_list(&list)
    } else if let Ok(vec) = value.extract::<Vec<Bound<'_, PyAny>>>() {
        frobnicate_vec(vec)
    } else {
        Err(PyTypeError::new_err("Cannot frobnicate that type."))
    }
}

This suboptimal as the FromPyObject<T> trait requires extract to have a Result<T, PyErr> return type. For native types like PyList, it faster to use downcast (which extract calls internally) when the error value is ignored. This avoids the costly conversion of a PyDowncastError to a PyErr required to fulfil the FromPyObject contract, i.e.

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::{exceptions::PyTypeError, types::PyList};
fn frobnicate_list<'py>(list: &Bound<'_, PyList>) -> PyResult<Bound<'py, PyAny>> { todo!() }
fn frobnicate_vec<'py>(vec: Vec<Bound<'py, PyAny>>) -> PyResult<Bound<'py, PyAny>> { todo!() }

#[pyfunction]
fn frobnicate<'py>(value: &Bound<'py, PyAny>) -> PyResult<Bound<'py, PyAny>> {
    // Use `downcast` instead of `extract` as turning `PyDowncastError` into `PyErr` is quite costly.
    if let Ok(list) = value.downcast::<PyList>() {
        frobnicate_list(list)
    } else if let Ok(vec) = value.extract::<Vec<Bound<'_, PyAny>>>() {
        frobnicate_vec(vec)
    } else {
        Err(PyTypeError::new_err("Cannot frobnicate that type."))
    }
}

Access to Bound implies access to GIL token

Calling Python::with_gil is effectively a no-op when the GIL is already held, but checking that this is the case still has a cost. If an existing GIL token can not be accessed, for example when implementing a pre-existing trait, but a GIL-bound reference is available, this cost can be avoided by exploiting that access to GIL-bound reference gives zero-cost access to a GIL token via Bound::py.

For example, instead of writing

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyList;

struct Foo(Py<PyList>);

struct FooBound<'py>(Bound<'py, PyList>);

impl PartialEq<Foo> for FooBound<'_> {
    fn eq(&self, other: &Foo) -> bool {
        Python::with_gil(|py| {
            let len = other.0.bind(py).len();
            self.0.len() == len
        })
    }
}

use the more efficient

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyList;
struct Foo(Py<PyList>);
struct FooBound<'py>(Bound<'py, PyList>);

impl PartialEq<Foo> for FooBound<'_> {
    fn eq(&self, other: &Foo) -> bool {
        // Access to `&Bound<'py, PyAny>` implies access to `Python<'py>`.
        let py = self.0.py();
        let len = other.0.bind(py).len();
        self.0.len() == len
    }
}

Advanced topics

FFI

PyO3 exposes much of Python's C API through the ffi module.

The C API is naturally unsafe and requires you to manage reference counts, errors and specific invariants yourself. Please refer to the C API Reference Manual and The Rustonomicon before using any function from that API.

Building and distribution

This chapter of the guide goes into detail on how to build and distribute projects using PyO3. The way to achieve this is very different depending on whether the project is a Python module implemented in Rust, or a Rust binary embedding Python. For both types of project there are also common problems such as the Python version to build for and the linker arguments to use.

The material in this chapter is intended for users who have already read the PyO3 README. It covers in turn the choices that can be made for Python modules and for Rust binaries. There is also a section at the end about cross-compiling projects using PyO3.

There is an additional sub-chapter dedicated to supporting multiple Python versions.

Configuring the Python version

PyO3 uses a build script (backed by the pyo3-build-config crate) to determine the Python version and set the correct linker arguments. By default it will attempt to use the following in order:

• Any active Python virtualenv.
• The python executable (if it's a Python 3 interpreter).
• The python3 executable.

You can override the Python interpreter by setting the PYO3_PYTHON environment variable, e.g. PYO3_PYTHON=python3.7, PYO3_PYTHON=/usr/bin/python3.9, or even a PyPy interpreter PYO3_PYTHON=pypy3.

Once the Python interpreter is located, pyo3-build-config executes it to query the information in the sysconfig module which is needed to configure the rest of the compilation.

To validate the configuration which PyO3 will use, you can run a compilation with the environment variable PYO3_PRINT_CONFIG=1 set. An example output of doing this is shown below:

$ PYO3_PRINT_CONFIG=1 cargo build
   Compiling pyo3 v0.14.1 (/home/david/dev/pyo3)
error: failed to run custom build command for `pyo3 v0.14.1 (/home/david/dev/pyo3)`

Caused by:
  process didn't exit successfully: `/home/david/dev/pyo3/target/debug/build/pyo3-7a8cf4fe22e959b7/build-script-build` (exit status: 101)
  --- stdout
  cargo:rerun-if-env-changed=PYO3_CROSS
  cargo:rerun-if-env-changed=PYO3_CROSS_LIB_DIR
  cargo:rerun-if-env-changed=PYO3_CROSS_PYTHON_VERSION
  cargo:rerun-if-env-changed=PYO3_PRINT_CONFIG

  -- PYO3_PRINT_CONFIG=1 is set, printing configuration and halting compile --
  implementation=CPython
  version=3.8
  shared=true
  abi3=false
  lib_name=python3.8
  lib_dir=/usr/lib
  executable=/usr/bin/python
  pointer_width=64
  build_flags=
  suppress_build_script_link_lines=false

The PYO3_ENVIRONMENT_SIGNATURE environment variable can be used to trigger rebuilds when its value changes, it has no other effect.

Advanced: config files

If you save the above output config from PYO3_PRINT_CONFIG to a file, it is possible to manually override the contents and feed it back into PyO3 using the PYO3_CONFIG_FILE env var.

If your build environment is unusual enough that PyO3's regular configuration detection doesn't work, using a config file like this will give you the flexibility to make PyO3 work for you. To see the full set of options supported, see the documentation for the InterpreterConfig struct.

Building Python extension modules

Python extension modules need to be compiled differently depending on the OS (and architecture) that they are being compiled for. As well as multiple OSes (and architectures), there are also many different Python versions which are actively supported. Packages uploaded to PyPI usually want to upload prebuilt "wheels" covering many OS/arch/version combinations so that users on all these different platforms don't have to compile the package themselves. Package vendors can opt-in to the "abi3" limited Python API which allows their wheels to be used on multiple Python versions, reducing the number of wheels they need to compile, but restricts the functionality they can use.

There are many ways to go about this: it is possible to use cargo to build the extension module (along with some manual work, which varies with OS). The PyO3 ecosystem has two packaging tools, maturin and setuptools-rust, which abstract over the OS difference and also support building wheels for PyPI upload.

PyO3 has some Cargo features to configure projects for building Python extension modules:

• The extension-module feature, which must be enabled when building Python extension modules.
• The abi3 feature and its version-specific abi3-pyXY companions, which are used to opt-in to the limited Python API in order to support multiple Python versions in a single wheel.

This section describes each of these packaging tools before describing how to build manually without them. It then proceeds with an explanation of the extension-module feature. Finally, there is a section describing PyO3's abi3 features.

Packaging tools

The PyO3 ecosystem has two main choices to abstract the process of developing Python extension modules:

• maturin is a command-line tool to build, package and upload Python modules. It makes opinionated choices about project layout meaning it needs very little configuration. This makes it a great choice for users who are building a Python extension from scratch and don't need flexibility.
• setuptools-rust is an add-on for setuptools which adds extra keyword arguments to the setup.py configuration file. It requires more configuration than maturin, however this gives additional flexibility for users adding Rust to an existing Python package that can't satisfy maturin's constraints.

Consult each project's documentation for full details on how to get started using them and how to upload wheels to PyPI. It should be noted that while maturin is able to build manylinux-compliant wheels out-of-the-box, setuptools-rust requires a bit more effort, relying on Docker for this purpose.

There are also maturin-starter and setuptools-rust-starter examples in the PyO3 repository.

Manual builds

To build a PyO3-based Python extension manually, start by running cargo build as normal in a library project which uses PyO3's extension-module feature and has the cdylib crate type.

Once built, symlink (or copy) and rename the shared library from Cargo's target/ directory to your desired output directory:

• on macOS, rename libyour_module.dylib to your_module.so.
• on Windows, rename libyour_module.dll to your_module.pyd.
• on Linux, rename libyour_module.so to your_module.so.

You can then open a Python shell in the output directory and you'll be able to run import your_module.

If you're packaging your library for redistribution, you should indicated the Python interpreter your library is compiled for by including the platform tag in its name. This prevents incompatible interpreters from trying to import your library. If you're compiling for PyPy you must include the platform tag, or PyPy will ignore the module.

Bazel builds

To use PyO3 with bazel one needs to manually configure PyO3, PyO3-ffi and PyO3-macros. In particular, one needs to make sure that it is compiled with the right python flags for the version you intend to use. For example see:

1. https://github.com/OliverFM/pytorch_with_gazelle -- for a minimal example of a repo that can use PyO3, PyTorch and Gazelle to generate python Build files.
2. https://github.com/TheButlah/rules_pyo3 -- which has more extensive support, but is outdated.

Platform tags

Rather than using just the .so or .pyd extension suggested above (depending on OS), you can prefix the shared library extension with a platform tag to indicate the interpreter it is compatible with. You can query your interpreter's platform tag from the sysconfig module. Some example outputs of this are seen below:

# CPython 3.10 on macOS
.cpython-310-darwin.so

# PyPy 7.3 (Python 3.8) on Linux
$ python -c 'import sysconfig; print(sysconfig.get_config_var("EXT_SUFFIX"))'
.pypy38-pp73-x86_64-linux-gnu.so

So, for example, a valid module library name on CPython 3.10 for macOS is your_module.cpython-310-darwin.so, and its equivalent when compiled for PyPy 7.3 on Linux would be your_module.pypy38-pp73-x86_64-linux-gnu.so.

See PEP 3149 for more background on platform tags.

macOS

On macOS, because the extension-module feature disables linking to libpython (see the next section), some additional linker arguments need to be set. maturin and setuptools-rust both pass these arguments for PyO3 automatically, but projects using manual builds will need to set these directly in order to support macOS.

The easiest way to set the correct linker arguments is to add a build.rs with the following content:

fn main() {
    pyo3_build_config::add_extension_module_link_args();
}

Remember to also add pyo3-build-config to the build-dependencies section in Cargo.toml.

An alternative to using pyo3-build-config is add the following to a cargo configuration file (e.g. .cargo/config.toml):

[target.x86_64-apple-darwin]
rustflags = [
  "-C", "link-arg=-undefined",
  "-C", "link-arg=dynamic_lookup",
]

[target.aarch64-apple-darwin]
rustflags = [
  "-C", "link-arg=-undefined",
  "-C", "link-arg=dynamic_lookup",
]

Using the MacOS system python3 (/usr/bin/python3, as opposed to python installed via homebrew, pyenv, nix, etc.) may result in runtime errors such as Library not loaded: @rpath/Python3.framework/Versions/3.8/Python3. These can be resolved with another addition to .cargo/config.toml:

[build]
rustflags = [
  "-C", "link-args=-Wl,-rpath,/Library/Developer/CommandLineTools/Library/Frameworks",
]

Alternatively, on rust >= 1.56, one can include in build.rs:

fn main() {
    println!(
        "cargo:rustc-link-arg=-Wl,-rpath,/Library/Developer/CommandLineTools/Library/Frameworks"
    );
}

For more discussion on and workarounds for MacOS linking problems see this issue.

Finally, don't forget that on MacOS the extension-module feature will cause cargo test to fail without the --no-default-features flag (see the FAQ).

The extension-module feature

PyO3's extension-module feature is used to disable linking to libpython on Unix targets.

This is necessary because by default PyO3 links to libpython. This makes binaries, tests, and examples "just work". However, Python extensions on Unix must not link to libpython for manylinux compliance.

The downside of not linking to libpython is that binaries, tests, and examples (which usually embed Python) will fail to build. If you have an extension module as well as other outputs in a single project, you need to use optional Cargo features to disable the extension-module when you're not building the extension module. See the FAQ for an example workaround.

Py_LIMITED_API/abi3

By default, Python extension modules can only be used with the same Python version they were compiled against. For example, an extension module built for Python 3.5 can't be imported in Python 3.8. PEP 384 introduced the idea of the limited Python API, which would have a stable ABI enabling extension modules built with it to be used against multiple Python versions. This is also known as abi3.

The advantage of building extension modules using the limited Python API is that package vendors only need to build and distribute a single copy (for each OS / architecture), and users can install it on all Python versions from the minimum version and up. The downside of this is that PyO3 can't use optimizations which rely on being compiled against a known exact Python version. It's up to you to decide whether this matters for your extension module. It's also possible to design your extension module such that you can distribute abi3 wheels but allow users compiling from source to benefit from additional optimizations - see the support for multiple python versions section of this guide, in particular the #[cfg(Py_LIMITED_API)] flag.

There are three steps involved in making use of abi3 when building Python packages as wheels:

1. Enable the abi3 feature in pyo3. This ensures pyo3 only calls Python C-API functions which are part of the stable API, and on Windows also ensures that the project links against the correct shared object (no special behavior is required on other platforms):

[dependencies]
pyo3 = { version = "0.21.2", features = ["abi3"] }

2. Ensure that the built shared objects are correctly marked as abi3. This is accomplished by telling your build system that you're using the limited API. maturin >= 0.9.0 and setuptools-rust >= 0.11.4 support abi3 wheels. See the corresponding PRs for more.

3. Ensure that the .whl is correctly marked as abi3. For projects using setuptools, this is accomplished by passing --py-limited-api=cp3x (where x is the minimum Python version supported by the wheel, e.g. --py-limited-api=cp35 for Python 3.5) to setup.py bdist_wheel.

Minimum Python version for abi3

Because a single abi3 wheel can be used with many different Python versions, PyO3 has feature flags abi3-py37, abi3-py38, abi3-py39 etc. to set the minimum required Python version for your abi3 wheel. For example, if you set the abi3-py37 feature, your extension wheel can be used on all Python 3 versions from Python 3.7 and up. maturin and setuptools-rust will give the wheel a name like my-extension-1.0-cp37-abi3-manylinux2020_x86_64.whl.

As your extension module may be run with multiple different Python versions you may occasionally find you need to check the Python version at runtime to customize behavior. See the relevant section of this guide on supporting multiple Python versions at runtime.

PyO3 is only able to link your extension module to abi3 version up to and including your host Python version. E.g., if you set abi3-py38 and try to compile the crate with a host of Python 3.7, the build will fail.

Note: If you set more that one of these abi3 version feature flags the lowest version always wins. For example, with both abi3-py37 and abi3-py38 set, PyO3 would build a wheel which supports Python 3.7 and up.

Building abi3 extensions without a Python interpreter

As an advanced feature, you can build PyO3 wheel without calling Python interpreter with the environment variable PYO3_NO_PYTHON set. Also, if the build host Python interpreter is not found or is too old or otherwise unusable, PyO3 will still attempt to compile abi3 extension modules after displaying a warning message. On Unix-like systems this works unconditionally; on Windows you must also set the RUSTFLAGS environment variable to contain -L native=/path/to/python/libs so that the linker can find python3.lib.

If the python3.dll import library is not available, an experimental generate-import-lib crate feature may be enabled, and the required library will be created and used by PyO3 automatically.

Note: MSVC targets require LLVM binutils (llvm-dlltool) to be available in PATH for the automatic import library generation feature to work.

Missing features

Due to limitations in the Python API, there are a few pyo3 features that do not work when compiling for abi3. These are:

• #[pyo3(text_signature = "...")] does not work on classes until Python 3.10 or greater.
• The dict and weakref options on classes are not supported until Python 3.9 or greater.
• The buffer API is not supported until Python 3.11 or greater.
• Optimizations which rely on knowledge of the exact Python version compiled against.

Embedding Python in Rust

If you want to embed the Python interpreter inside a Rust program, there are two modes in which this can be done: dynamically and statically. We'll cover each of these modes in the following sections. Each of them affect how you must distribute your program. Instead of learning how to do this yourself, you might want to consider using a project like PyOxidizer to ship your application and all of its dependencies in a single file.

PyO3 automatically switches between the two linking modes depending on whether the Python distribution you have configured PyO3 to use (see above) contains a shared library or a static library. The static library is most often seen in Python distributions compiled from source without the --enable-shared configuration option. For example, this is the default for pyenv on macOS.

Dynamically embedding the Python interpreter

Embedding the Python interpreter dynamically is much easier than doing so statically. This is done by linking your program against a Python shared library (such as libpython.3.9.so on UNIX, or python39.dll on Windows). The implementation of the Python interpreter resides inside the shared library. This means that when the OS runs your Rust program it also needs to be able to find the Python shared library.

This mode of embedding works well for Rust tests which need access to the Python interpreter. It is also great for Rust software which is installed inside a Python virtualenv, because the virtualenv sets up appropriate environment variables to locate the correct Python shared library.

For distributing your program to non-technical users, you will have to consider including the Python shared library in your distribution as well as setting up wrapper scripts to set the right environment variables (such as LD_LIBRARY_PATH on UNIX, or PATH on Windows).

Note that PyPy cannot be embedded in Rust (or any other software). Support for this is tracked on the PyPy issue tracker.

Statically embedding the Python interpreter

Embedding the Python interpreter statically means including the contents of a Python static library directly inside your Rust binary. This means that to distribute your program you only need to ship your binary file: it contains the Python interpreter inside the binary!

On Windows static linking is almost never done, so Python distributions don't usually include a static library. The information below applies only to UNIX.

The Python static library is usually called libpython.a.

Static linking has a lot of complications, listed below. For these reasons PyO3 does not yet have first-class support for this embedding mode. See issue 416 on PyO3's GitHub for more information and to discuss any issues you encounter.

The auto-initialize feature is deliberately disabled when embedding the interpreter statically because this is often unintentionally done by new users to PyO3 running test programs. Trying out PyO3 is much easier using dynamic embedding.

The known complications are:

• To import compiled extension modules (such as other Rust extension modules, or those written in C), your binary must have the correct linker flags set during compilation to export the original contents of libpython.a so that extensions can use them (e.g. -Wl,--export-dynamic).

• The C compiler and flags which were used to create libpython.a must be compatible with your Rust compiler and flags, else you will experience compilation failures.

  Significantly different compiler versions may see errors like this:

  lto1: fatal error: bytecode stream in file 'rust-numpy/target/release/deps/libpyo3-6a7fb2ed970dbf26.rlib' generated with LTO version 6.0 instead of the expected 6.2
  

  Mismatching flags may lead to errors like this:

  /usr/bin/ld: /usr/lib/gcc/x86_64-linux-gnu/9/../../../x86_64-linux-gnu/libpython3.9.a(zlibmodule.o): relocation R_X86_64_32 against `.data' can not be used when making a PIE object; recompile with -fPIE
  

If you encounter these or other complications when linking the interpreter statically, discuss them on issue 416 on PyO3's GitHub. It is hoped that eventually that discussion will contain enough information and solutions that PyO3 can offer first-class support for static embedding.

Import your module when embedding the Python interpreter

When you run your Rust binary with an embedded interpreter, any #[pymodule] created modules won't be accessible to import unless added to a table called PyImport_Inittab before the embedded interpreter is initialized. This will cause Python statements in your embedded interpreter such as import your_new_module to fail. You can call the macro append_to_inittab with your module before initializing the Python interpreter to add the module function into that table. (The Python interpreter will be initialized by calling prepare_freethreaded_python, with_embedded_python_interpreter, or Python::with_gil with the auto-initialize feature enabled.)

Cross Compiling

Thanks to Rust's great cross-compilation support, cross-compiling using PyO3 is relatively straightforward. To get started, you'll need a few pieces of software:

• A toolchain for your target.
• The appropriate options in your Cargo .config for the platform you're targeting and the toolchain you are using.
• A Python interpreter that's already been compiled for your target (optional when building "abi3" extension modules).
• A Python interpreter that is built for your host and available through the PATH or setting the PYO3_PYTHON variable (optional when building "abi3" extension modules).

After you've obtained the above, you can build a cross-compiled PyO3 module by using Cargo's --target flag. PyO3's build script will detect that you are attempting a cross-compile based on your host machine and the desired target.

When cross-compiling, PyO3's build script cannot execute the target Python interpreter to query the configuration, so there are a few additional environment variables you may need to set:

• PYO3_CROSS: If present this variable forces PyO3 to configure as a cross-compilation.
• PYO3_CROSS_LIB_DIR: This variable can be set to the directory containing the target's libpython DSO and the associated _sysconfigdata*.py file for Unix-like targets, or the Python DLL import libraries for the Windows target. This variable is only needed when the output binary must link to libpython explicitly (e.g. when targeting Windows and Android or embedding a Python interpreter), or when it is absolutely required to get the interpreter configuration from _sysconfigdata*.py.
• PYO3_CROSS_PYTHON_VERSION: Major and minor version (e.g. 3.9) of the target Python installation. This variable is only needed if PyO3 cannot determine the version to target from abi3-py3* features, or if PYO3_CROSS_LIB_DIR is not set, or if there are multiple versions of Python present in PYO3_CROSS_LIB_DIR.
• PYO3_CROSS_PYTHON_IMPLEMENTATION: Python implementation name ("CPython" or "PyPy") of the target Python installation. CPython is assumed by default when this variable is not set, unless PYO3_CROSS_LIB_DIR is set for a Unix-like target and PyO3 can get the interpreter configuration from _sysconfigdata*.py.

An experimental pyo3 crate feature generate-import-lib enables the user to cross-compile extension modules for Windows targets without setting the PYO3_CROSS_LIB_DIR environment variable or providing any Windows Python library files. It uses an external python3-dll-a crate to generate import libraries for the Python DLL for MinGW-w64 and MSVC compile targets. python3-dll-a uses the binutils dlltool program to generate DLL import libraries for MinGW-w64 targets. It is possible to override the default dlltool command name for the cross target by setting PYO3_MINGW_DLLTOOL environment variable. Note: MSVC targets require LLVM binutils or MSVC build tools to be available on the host system. More specifically, python3-dll-a requires llvm-dlltool or lib.exe executable to be present in PATH when targeting *-pc-windows-msvc. The Zig compiler executable can be used in place of llvm-dlltool when the ZIG_COMMAND environment variable is set to the installed Zig program name ("zig" or "python -m ziglang").

An example might look like the following (assuming your target's sysroot is at /home/pyo3/cross/sysroot and that your target is armv7):

export PYO3_CROSS_LIB_DIR="/home/pyo3/cross/sysroot/usr/lib"

cargo build --target armv7-unknown-linux-gnueabihf

If there are multiple python versions at the cross lib directory and you cannot set a more precise location to include both the libpython DSO and _sysconfigdata*.py files, you can set the required version:

export PYO3_CROSS_PYTHON_VERSION=3.8
export PYO3_CROSS_LIB_DIR="/home/pyo3/cross/sysroot/usr/lib"

cargo build --target armv7-unknown-linux-gnueabihf

Or another example with the same sys root but building for Windows:

export PYO3_CROSS_PYTHON_VERSION=3.9
export PYO3_CROSS_LIB_DIR="/home/pyo3/cross/sysroot/usr/lib"

cargo build --target x86_64-pc-windows-gnu

Any of the abi3-py3* features can be enabled instead of setting PYO3_CROSS_PYTHON_VERSION in the above examples.

PYO3_CROSS_LIB_DIR can often be omitted when cross compiling extension modules for Unix and macOS targets, or when cross compiling extension modules for Windows and the experimental generate-import-lib crate feature is enabled.

The following resources may also be useful for cross-compiling:

• github.com/japaric/rust-cross is a primer on cross compiling Rust.
• github.com/rust-embedded/cross uses Docker to make Rust cross-compilation easier.

Supporting multiple Python versions

PyO3 supports all actively-supported Python 3 and PyPy versions. As much as possible, this is done internally to PyO3 so that your crate's code does not need to adapt to the differences between each version. However, as Python features grow and change between versions, PyO3 cannot a completely identical API for every Python version. This may require you to add conditional compilation to your crate or runtime checks for the Python version.

This section of the guide first introduces the pyo3-build-config crate, which you can use as a build-dependency to add additional #[cfg] flags which allow you to support multiple Python versions at compile-time.

Second, we'll show how to check the Python version at runtime. This can be useful when building for multiple versions with the abi3 feature, where the Python API compiled against is not always the same as the one in use.

Conditional compilation for different Python versions

The pyo3-build-config exposes multiple #[cfg] flags which can be used to conditionally compile code for a given Python version. PyO3 itself depends on this crate, so by using it you can be sure that you are configured correctly for the Python version PyO3 is building against.

This allows us to write code like the following

#[cfg(Py_3_7)]
fn function_only_supported_on_python_3_7_and_up() {}

#[cfg(not(Py_3_8))]
fn function_only_supported_before_python_3_8() {}

#[cfg(not(Py_LIMITED_API))]
fn function_incompatible_with_abi3_feature() {}

The following sections first show how to add these #[cfg] flags to your build process, and then cover some common patterns flags in a little more detail.

To see a full reference of all the #[cfg] flags provided, see the pyo3-build-cfg docs.

Using pyo3-build-config

You can use the #[cfg] flags in just two steps:

1. Add pyo3-build-config with the resolve-config feature enabled to your crate's build dependencies in Cargo.toml:

   [build-dependencies]
   pyo3-build-config = { version = "0.21.2", features = ["resolve-config"] }
   

2. Add a build.rs file to your crate with the following contents:

   fn main() {
       // If you have an existing build.rs file, just add this line to it.
       pyo3_build_config::use_pyo3_cfgs();
   }

After these steps you are ready to annotate your code!

Common usages of pyo3-build-cfg flags

The #[cfg] flags added by pyo3-build-cfg can be combined with all of Rust's logic in the #[cfg] attribute to create very precise conditional code generation. The following are some common patterns implemented using these flags:

#[cfg(Py_3_7)]

This #[cfg] marks code that will only be present on Python 3.7 and upwards. There are similar options Py_3_8, Py_3_9, Py_3_10 and so on for each minor version.

#[cfg(not(Py_3_7))]

This #[cfg] marks code that will only be present on Python versions before (but not including) Python 3.7.

#[cfg(not(Py_LIMITED_API))]

This #[cfg] marks code that is only available when building for the unlimited Python API (i.e. PyO3's abi3 feature is not enabled). This might be useful if you want to ship your extension module as an abi3 wheel and also allow users to compile it from source to make use of optimizations only possible with the unlimited API.

#[cfg(any(Py_3_9, not(Py_LIMITED_API)))]

This #[cfg] marks code which is available when running Python 3.9 or newer, or when using the unlimited API with an older Python version. Patterns like this are commonly seen on Python APIs which were added to the limited Python API in a specific minor version.

#[cfg(PyPy)]

This #[cfg] marks code which is running on PyPy.

Checking the Python version at runtime

When building with PyO3's abi3 feature, your extension module will be compiled against a specific minimum version of Python, but may be running on newer Python versions.

For example with PyO3's abi3-py38 feature, your extension will be compiled as if it were for Python 3.8. If you were using pyo3-build-config, #[cfg(Py_3_8)] would be present. Your user could freely install and run your abi3 extension on Python 3.9.

There's no way to detect your user doing that at compile time, so instead you need to fall back to runtime checks.

PyO3 provides the APIs Python::version() and Python::version_info() to query the running Python version. This allows you to do the following, for example:

use pyo3::Python;

Python::with_gil(|py| {
    // PyO3 supports Python 3.7 and up.
    assert!(py.version_info() >= (3, 7));
    assert!(py.version_info() >= (3, 7, 0));
});

The PyO3 ecosystem

This portion of the guide is dedicated to crates which are external to the main PyO3 project and provide additional functionality you might find useful.

Because these projects evolve independently of the PyO3 repository the content of these articles may fall out of date over time; please file issues on the PyO3 GitHub to alert maintainers when this is the case.

Logging

It is desirable if both the Python and Rust parts of the application end up logging using the same configuration into the same place.

This section of the guide briefly discusses how to connect the two languages' logging ecosystems together. The recommended way for Python extension modules is to configure Rust's logger to send log messages to Python using the pyo3-log crate. For users who want to do the opposite and send Python log messages to Rust, see the note at the end of this guide.

Using pyo3-log to send Rust log messages to Python

The pyo3-log crate allows sending the messages from the Rust side to Python's logging system. This is mostly suitable for writing native extensions for Python programs.

Use pyo3_log::init to install the logger in its default configuration. It's also possible to tweak its configuration (mostly to tune its performance).

use log::info;
use pyo3::prelude::*;

#[pyfunction]
fn log_something() {
    // This will use the logger installed in `my_module` to send the `info`
    // message to the Python logging facilities.
    info!("Something!");
}

#[pymodule]
fn my_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    // A good place to install the Rust -> Python logger.
    pyo3_log::init();

    m.add_function(wrap_pyfunction!(log_something, m)?)?;
    Ok(())
}

Then it is up to the Python side to actually output the messages somewhere.

import logging
import my_module

FORMAT = '%(levelname)s %(name)s %(asctime)-15s %(filename)s:%(lineno)d %(message)s'
logging.basicConfig(format=FORMAT)
logging.getLogger().setLevel(logging.INFO)
my_module.log_something()

It is important to initialize the Python loggers first, before calling any Rust functions that may log. This limitation can be worked around if it is not possible to satisfy, read the documentation about caching.

The Python to Rust direction

To have python logs be handled by Rust, one need only register a rust function to handle logs emitted from the core python logging module.

This has been implemented within the pyo3-pylogger crate.

use log::{info, warn};
use pyo3::prelude::*;

fn main() -> PyResult<()> {
    // register the host handler with python logger, providing a logger target
    // set the name here to something appropriate for your application
    pyo3_pylogger::register("example_application_py_logger");

    // initialize up a logger
    env_logger::Builder::from_env(env_logger::Env::default().default_filter_or("trace")).init();

    // Log some messages from Rust.
    info!("Just some normal information!");
    warn!("Something spooky happened!");

    // Log some messages from Python
    Python::with_gil(|py| {
        py.run(
            "
import logging
logging.error('Something bad happened')
",
            None,
            None,
        )
    })
}

Using async and await

async/await support is currently being integrated in PyO3. See the dedicated documentation

If you are working with a Python library that makes use of async functions or wish to provide Python bindings for an async Rust library, pyo3-asyncio likely has the tools you need. It provides conversions between async functions in both Python and Rust and was designed with first-class support for popular Rust runtimes such as tokio and async-std. In addition, all async Python code runs on the default asyncio event loop, so pyo3-asyncio should work just fine with existing Python libraries.

In the following sections, we'll give a general overview of pyo3-asyncio explaining how to call async Python functions with PyO3, how to call async Rust functions from Python, and how to configure your codebase to manage the runtimes of both.

Quickstart

Here are some examples to get you started right away! A more detailed breakdown of the concepts in these examples can be found in the following sections.

Rust Applications

Here we initialize the runtime, import Python's asyncio library and run the given future to completion using Python's default EventLoop and async-std. Inside the future, we convert asyncio sleep into a Rust future and await it.

# Cargo.toml dependencies
[dependencies]
pyo3 = { version = "0.14" }
pyo3-asyncio = { version = "0.14", features = ["attributes", "async-std-runtime"] }
async-std = "1.9"

//! main.rs

use pyo3::prelude::*;

#[pyo3_asyncio::async_std::main]
async fn main() -> PyResult<()> {
    let fut = Python::with_gil(|py| {
        let asyncio = py.import("asyncio")?;
        // convert asyncio.sleep into a Rust Future
        pyo3_asyncio::async_std::into_future(asyncio.call_method1("sleep", (1.into_py(py),))?)
    })?;

    fut.await?;

    Ok(())
}

The same application can be written to use tokio instead using the #[pyo3_asyncio::tokio::main] attribute.

# Cargo.toml dependencies
[dependencies]
pyo3 = { version = "0.14" }
pyo3-asyncio = { version = "0.14", features = ["attributes", "tokio-runtime"] }
tokio = "1.4"

//! main.rs

use pyo3::prelude::*;

#[pyo3_asyncio::tokio::main]
async fn main() -> PyResult<()> {
    let fut = Python::with_gil(|py| {
        let asyncio = py.import("asyncio")?;
        // convert asyncio.sleep into a Rust Future
        pyo3_asyncio::tokio::into_future(asyncio.call_method1("sleep", (1.into_py(py),))?)
    })?;

    fut.await?;

    Ok(())
}

More details on the usage of this library can be found in the API docs and the primer below.

PyO3 Native Rust Modules

PyO3 Asyncio can also be used to write native modules with async functions.

Add the [lib] section to Cargo.toml to make your library a cdylib that Python can import.

[lib]
name = "my_async_module"
crate-type = ["cdylib"]

Make your project depend on pyo3 with the extension-module feature enabled and select your pyo3-asyncio runtime:

For async-std:

[dependencies]
pyo3 = { version = "0.14", features = ["extension-module"] }
pyo3-asyncio = { version = "0.14", features = ["async-std-runtime"] }
async-std = "1.9"

For tokio:

[dependencies]
pyo3 = { version = "0.14", features = ["extension-module"] }
pyo3-asyncio = { version = "0.14", features = ["tokio-runtime"] }
tokio = "1.4"

Export an async function that makes use of async-std:

//! lib.rs

use pyo3::{prelude::*, wrap_pyfunction};

#[pyfunction]
fn rust_sleep(py: Python<'_>) -> PyResult<&Bound<'_, PyAny>> {
    pyo3_asyncio::async_std::future_into_py(py, async {
        async_std::task::sleep(std::time::Duration::from_secs(1)).await;
        Ok(Python::with_gil(|py| py.None()))
    })
}

#[pymodule]
fn my_async_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(rust_sleep, m)?)?;

    Ok(())
}

If you want to use tokio instead, here's what your module should look like:

//! lib.rs

use pyo3::{prelude::*, wrap_pyfunction};

#[pyfunction]
fn rust_sleep(py: Python<'_>) -> PyResult<&Bound<'_, PyAny>>> {
    pyo3_asyncio::tokio::future_into_py(py, async {
        tokio::time::sleep(std::time::Duration::from_secs(1)).await;
        Ok(Python::with_gil(|py| py.None()))
    })
}

#[pymodule]
fn my_async_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(rust_sleep, m)?)?;
    Ok(())
}

You can build your module with maturin (see the Using Rust in Python section in the PyO3 guide for setup instructions). After that you should be able to run the Python REPL to try it out.

maturin develop && python3
🔗 Found pyo3 bindings
🐍 Found CPython 3.8 at python3
    Finished dev [unoptimized + debuginfo] target(s) in 0.04s
Python 3.8.5 (default, Jan 27 2021, 15:41:15)
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import asyncio
>>>
>>> from my_async_module import rust_sleep
>>>
>>> async def main():
>>>     await rust_sleep()
>>>
>>> # should sleep for 1s
>>> asyncio.run(main())
>>>

Awaiting an Async Python Function in Rust

Let's take a look at a dead simple async Python function:

# Sleep for 1 second
async def py_sleep():
    await asyncio.sleep(1)

Async functions in Python are simply functions that return a coroutine object. For our purposes, we really don't need to know much about these coroutine objects. The key factor here is that calling an async function is just like calling a regular function, the only difference is that we have to do something special with the object that it returns.

Normally in Python, that something special is the await keyword, but in order to await this coroutine in Rust, we first need to convert it into Rust's version of a coroutine: a Future. That's where pyo3-asyncio comes in. pyo3_asyncio::async_std::into_future performs this conversion for us.

The following example uses into_future to call the py_sleep function shown above and then await the coroutine object returned from the call:

use pyo3::prelude::*;

#[pyo3_asyncio::tokio::main]
async fn main() -> PyResult<()> {
    let future = Python::with_gil(|py| -> PyResult<_> {
        // import the module containing the py_sleep function
        let example = py.import("example")?;

        // calling the py_sleep method like a normal function
        // returns a coroutine
        let coroutine = example.call_method0("py_sleep")?;

        // convert the coroutine into a Rust future using the
        // tokio runtime
        pyo3_asyncio::tokio::into_future(coroutine)
    })?;

    // await the future
    future.await?;

    Ok(())
}

Alternatively, the below example shows how to write a #[pyfunction] which uses into_future to receive and await a coroutine argument:

#[pyfunction]
fn await_coro(coro: &Bound<'_, PyAny>>) -> PyResult<()> {
    // convert the coroutine into a Rust future using the
    // async_std runtime
    let f = pyo3_asyncio::async_std::into_future(coro)?;

    pyo3_asyncio::async_std::run_until_complete(coro.py(), async move {
        // await the future
        f.await?;
        Ok(())
    })
}

This could be called from Python as:

import asyncio

async def py_sleep():
    asyncio.sleep(1)

await_coro(py_sleep())

If for you wanted to pass a callable function to the #[pyfunction] instead, (i.e. the last line becomes await_coro(py_sleep)), then the above example needs to be tweaked to first call the callable to get the coroutine:

#[pyfunction]
fn await_coro(callable: &Bound<'_, PyAny>>) -> PyResult<()> {
    // get the coroutine by calling the callable
    let coro = callable.call0()?;

    // convert the coroutine into a Rust future using the
    // async_std runtime
    let f = pyo3_asyncio::async_std::into_future(coro)?;

    pyo3_asyncio::async_std::run_until_complete(coro.py(), async move {
        // await the future
        f.await?;
        Ok(())
    })
}

This can be particularly helpful where you need to repeatedly create and await a coroutine. Trying to await the same coroutine multiple times will raise an error:

RuntimeError: cannot reuse already awaited coroutine

If you're interested in learning more about coroutines and awaitables in general, check out the Python 3 asyncio docs for more information.

Awaiting a Rust Future in Python

Here we have the same async function as before written in Rust using the async-std runtime:

/// Sleep for 1 second
async fn rust_sleep() {
    async_std::task::sleep(std::time::Duration::from_secs(1)).await;
}

Similar to Python, Rust's async functions also return a special object called a Future:

let future = rust_sleep();

We can convert this Future object into Python to make it awaitable. This tells Python that you can use the await keyword with it. In order to do this, we'll call pyo3_asyncio::async_std::future_into_py:

use pyo3::prelude::*;

async fn rust_sleep() {
    async_std::task::sleep(std::time::Duration::from_secs(1)).await;
}

#[pyfunction]
fn call_rust_sleep(py: Python<'_>) -> PyResult<&Bound<'_, PyAny>>> {
    pyo3_asyncio::async_std::future_into_py(py, async move {
        rust_sleep().await;
        Ok(Python::with_gil(|py| py.None()))
    })
}

In Python, we can call this pyo3 function just like any other async function:

from example import call_rust_sleep

async def rust_sleep():
    await call_rust_sleep()

Managing Event Loops

Python's event loop requires some special treatment, especially regarding the main thread. Some of Python's asyncio features, like proper signal handling, require control over the main thread, which doesn't always play well with Rust.

Luckily, Rust's event loops are pretty flexible and don't need control over the main thread, so in pyo3-asyncio, we decided the best way to handle Rust/Python interop was to just surrender the main thread to Python and run Rust's event loops in the background. Unfortunately, since most event loop implementations prefer control over the main thread, this can still make some things awkward.

PyO3 Asyncio Initialization

Because Python needs to control the main thread, we can't use the convenient proc macros from Rust runtimes to handle the main function or #[test] functions. Instead, the initialization for PyO3 has to be done from the main function and the main thread must block on pyo3_asyncio::async_std::run_until_complete.

Because we have to block on one of those functions, we can't use #[async_std::main] or #[tokio::main] since it's not a good idea to make long blocking calls during an async function.

Internally, these #[main] proc macros are expanded to something like this:

fn main() {
    // your async main fn
    async fn _main_impl() { /* ... */ }
    Runtime::new().block_on(_main_impl());
}

Making a long blocking call inside the Future that's being driven by block_on prevents that thread from doing anything else and can spell trouble for some runtimes (also this will actually deadlock a single-threaded runtime!). Many runtimes have some sort of spawn_blocking mechanism that can avoid this problem, but again that's not something we can use here since we need it to block on the main thread.

For this reason, pyo3-asyncio provides its own set of proc macros to provide you with this initialization. These macros are intended to mirror the initialization of async-std and tokio while also satisfying the Python runtime's needs.

Here's a full example of PyO3 initialization with the async-std runtime:

use pyo3::prelude::*;

#[pyo3_asyncio::async_std::main]
async fn main() -> PyResult<()> {
    // PyO3 is initialized - Ready to go

    let fut = Python::with_gil(|py| -> PyResult<_> {
        let asyncio = py.import("asyncio")?;

        // convert asyncio.sleep into a Rust Future
        pyo3_asyncio::async_std::into_future(
            asyncio.call_method1("sleep", (1.into_py(py),))?
        )
    })?;

    fut.await?;

    Ok(())
}

A Note About asyncio.run

In Python 3.7+, the recommended way to run a top-level coroutine with asyncio is with asyncio.run. In v0.13 we recommended against using this function due to initialization issues, but in v0.14 it's perfectly valid to use this function... with a caveat.

Since our Rust <--> Python conversions require a reference to the Python event loop, this poses a problem. Imagine we have a PyO3 Asyncio module that defines a rust_sleep function like in previous examples. You might rightfully assume that you can call pass this directly into asyncio.run like this:

import asyncio

from my_async_module import rust_sleep

asyncio.run(rust_sleep())

You might be surprised to find out that this throws an error:

Traceback (most recent call last):
  File "example.py", line 5, in <module>
    asyncio.run(rust_sleep())
RuntimeError: no running event loop

What's happening here is that we are calling rust_sleep before the future is actually running on the event loop created by asyncio.run. This is counter-intuitive, but expected behaviour, and unfortunately there doesn't seem to be a good way of solving this problem within PyO3 Asyncio itself.

However, we can make this example work with a simple workaround:

import asyncio

from my_async_module import rust_sleep

# Calling main will just construct the coroutine that later calls rust_sleep.
# - This ensures that rust_sleep will be called when the event loop is running,
#   not before.
async def main():
    await rust_sleep()

# Run the main() coroutine at the top-level instead
asyncio.run(main())

Non-standard Python Event Loops

Python allows you to use alternatives to the default asyncio event loop. One popular alternative is uvloop. In v0.13 using non-standard event loops was a bit of an ordeal, but in v0.14 it's trivial.

Using uvloop in a PyO3 Asyncio Native Extensions

# Cargo.toml

[lib]
name = "my_async_module"
crate-type = ["cdylib"]

[dependencies]
pyo3 = { version = "0.14", features = ["extension-module"] }
pyo3-asyncio = { version = "0.14", features = ["tokio-runtime"] }
async-std = "1.9"
tokio = "1.4"

//! lib.rs

use pyo3::{prelude::*, wrap_pyfunction};

#[pyfunction]
fn rust_sleep(py: Python<'_>) -> PyResult<&Bound<'_, PyAny>>> {
    pyo3_asyncio::tokio::future_into_py(py, async {
        tokio::time::sleep(std::time::Duration::from_secs(1)).await;
        Ok(Python::with_gil(|py| py.None()))
    })
}

#[pymodule]
fn my_async_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_function(wrap_pyfunction!(rust_sleep, m)?)?;

    Ok(())
}

$ maturin develop && python3
🔗 Found pyo3 bindings
🐍 Found CPython 3.8 at python3
    Finished dev [unoptimized + debuginfo] target(s) in 0.04s
Python 3.8.8 (default, Apr 13 2021, 19:58:26)
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import asyncio
>>> import uvloop
>>>
>>> import my_async_module
>>>
>>> uvloop.install()
>>>
>>> async def main():
...     await my_async_module.rust_sleep()
...
>>> asyncio.run(main())
>>>

Using uvloop in Rust Applications

Using uvloop in Rust applications is a bit trickier, but it's still possible with relatively few modifications.

Unfortunately, we can't make use of the #[pyo3_asyncio::<runtime>::main] attribute with non-standard event loops. This is because the #[pyo3_asyncio::<runtime>::main] proc macro has to interact with the Python event loop before we can install the uvloop policy.

[dependencies]
async-std = "1.9"
pyo3 = "0.14"
pyo3-asyncio = { version = "0.14", features = ["async-std-runtime"] }

//! main.rs

use pyo3::{prelude::*, types::PyType};

fn main() -> PyResult<()> {
    pyo3::prepare_freethreaded_python();

    Python::with_gil(|py| {
        let uvloop = py.import("uvloop")?;
        uvloop.call_method0("install")?;

        // store a reference for the assertion
        let uvloop = PyObject::from(uvloop);

        pyo3_asyncio::async_std::run(py, async move {
            // verify that we are on a uvloop.Loop
            Python::with_gil(|py| -> PyResult<()> {
                assert!(pyo3_asyncio::async_std::get_current_loop(py)?.is_instance(
                    uvloop
                        .as_ref(py)
                        .getattr("Loop")?
                )?);
                Ok(())
            })?;

            async_std::task::sleep(std::time::Duration::from_secs(1)).await;

            Ok(())
        })
    })
}

Additional Information

• Managing event loop references can be tricky with pyo3-asyncio. See Event Loop References in the API docs to get a better intuition for how event loop references are managed in this library.
• Testing pyo3-asyncio libraries and applications requires a custom test harness since Python requires control over the main thread. You can find a testing guide in the API docs for the testing module

Frequently Asked Questions and troubleshooting

Sorry that you're having trouble using PyO3. If you can't find the answer to your problem in the list below, you can also reach out for help on GitHub Discussions and on Discord.

I'm experiencing deadlocks using PyO3 with lazy_static or once_cell!

lazy_static and once_cell::sync both use locks to ensure that initialization is performed only by a single thread. Because the Python GIL is an additional lock this can lead to deadlocks in the following way:

1. A thread (thread A) which has acquired the Python GIL starts initialization of a lazy_static value.
2. The initialization code calls some Python API which temporarily releases the GIL e.g. Python::import.
3. Another thread (thread B) acquires the Python GIL and attempts to access the same lazy_static value.
4. Thread B is blocked, because it waits for lazy_static's initialization to lock to release.
5. Thread A is blocked, because it waits to re-acquire the GIL which thread B still holds.
6. Deadlock.

PyO3 provides a struct GILOnceCell which works equivalently to OnceCell but relies solely on the Python GIL for thread safety. This means it can be used in place of lazy_static or once_cell where you are experiencing the deadlock described above. See the documentation for GILOnceCell for an example how to use it.

I can't run cargo test; or I can't build in a Cargo workspace: I'm having linker issues like "Symbol not found" or "Undefined reference to _PyExc_SystemError"!

Currently, #340 causes cargo test to fail with linking errors when the extension-module feature is activated. Linking errors can also happen when building in a cargo workspace where a different crate also uses PyO3 (see #2521). For now, there are three ways we can work around these issues.

1. Make the extension-module feature optional. Build with maturin develop --features "extension-module"

[dependencies.pyo3]
version = "0.21.2"

[features]
extension-module = ["pyo3/extension-module"]

2. Make the extension-module feature optional and default. Run tests with cargo test --no-default-features:

[dependencies.pyo3]
version = "0.21.2"

[features]
extension-module = ["pyo3/extension-module"]
default = ["extension-module"]

3. If you are using a pyproject.toml file to control maturin settings, add the following section:

[tool.maturin]
features = ["pyo3/extension-module"]
# Or for maturin 0.12:
# cargo-extra-args = ["--features", "pyo3/extension-module"]

I can't run cargo test: my crate cannot be found for tests in tests/ directory!

The Rust book suggests to put integration tests inside a tests/ directory.

For a PyO3 extension-module project where the crate-type is set to "cdylib" in your Cargo.toml, the compiler won't be able to find your crate and will display errors such as E0432 or E0463:

error[E0432]: unresolved import `my_crate`
 --> tests/test_my_crate.rs:1:5
  |
1 | use my_crate;
  |     ^^^^^^^^^^^^ no external crate `my_crate`

The best solution is to make your crate types include both rlib and cdylib:

# Cargo.toml
[lib]
crate-type = ["cdylib", "rlib"]

Ctrl-C doesn't do anything while my Rust code is executing!

This is because Ctrl-C raises a SIGINT signal, which is handled by the calling Python process by simply setting a flag to action upon later. This flag isn't checked while Rust code called from Python is executing, only once control returns to the Python interpreter.

You can give the Python interpreter a chance to process the signal properly by calling Python::check_signals. It's good practice to call this function regularly if you have a long-running Rust function so that your users can cancel it.

#[pyo3(get)] clones my field!

You may have a nested struct similar to this:

use pyo3::prelude::*;
#[pyclass]
#[derive(Clone)]
struct Inner {/* fields omitted */}

#[pyclass]
struct Outer {
    #[pyo3(get)]
    inner: Inner,
}

#[pymethods]
impl Outer {
    #[new]
    fn __new__() -> Self {
        Self { inner: Inner {} }
    }
}

When Python code accesses Outer's field, PyO3 will return a new object on every access (note that their addresses are different):

outer = Outer()

a = outer.inner
b = outer.inner

assert a is b, f"a: {a}\nb: {b}"

AssertionError: a: <builtins.Inner object at 0x00000238FFB9C7B0>
b: <builtins.Inner object at 0x00000238FFB9C830>

This can be especially confusing if the field is mutable, as getting the field and then mutating it won't persist - you'll just get a fresh clone of the original on the next access. Unfortunately Python and Rust don't agree about ownership - if PyO3 gave out references to (possibly) temporary Rust objects to Python code, Python code could then keep that reference alive indefinitely. Therefore returning Rust objects requires cloning.

If you don't want that cloning to happen, a workaround is to allocate the field on the Python heap and store a reference to that, by using Py<...>:

use pyo3::prelude::*;
#[pyclass]
#[derive(Clone)]
struct Inner {/* fields omitted */}

#[pyclass]
struct Outer {
    #[pyo3(get)]
    inner: Py<Inner>,
}

#[pymethods]
impl Outer {
    #[new]
    fn __new__(py: Python<'_>) -> PyResult<Self> {
        Ok(Self {
            inner: Py::new(py, Inner {})?,
        })
    }
}

This time a and b are the same object:

outer = Outer()

a = outer.inner
b = outer.inner

assert a is b, f"a: {a}\nb: {b}"
print(f"a: {a}\nb: {b}")

a: <builtins.Inner object at 0x0000020044FCC670>
b: <builtins.Inner object at 0x0000020044FCC670>

The downside to this approach is that any Rust code working on the Outer struct now has to acquire the GIL to do anything with its field.

I want to use the pyo3 crate re-exported from dependency but the proc-macros fail!

All PyO3 proc-macros (#[pyclass], #[pyfunction], #[derive(FromPyObject)] and so on) expect the pyo3 crate to be available under that name in your crate root, which is the normal situation when pyo3 is a direct dependency of your crate.

However, when the dependency is renamed, or your crate only indirectly depends on pyo3, you need to let the macro code know where to find the crate. This is done with the crate attribute:

use pyo3::prelude::*;
pub extern crate pyo3;
mod reexported { pub use ::pyo3; }
#[pyclass]
#[pyo3(crate = "reexported::pyo3")]
struct MyClass;

I'm trying to call Python from Rust but I get STATUS_DLL_NOT_FOUND or STATUS_ENTRYPOINT_NOT_FOUND!

This happens on Windows when linking to the python DLL fails or the wrong one is linked. The Python DLL on Windows will usually be called something like:

• python3X.dll for Python 3.X, e.g. python310.dll for Python 3.10
• python3.dll when using PyO3's abi3 feature

The DLL needs to be locatable using the Windows DLL search order. Some ways to achieve this are:

• Put the Python DLL in the same folder as your build artifacts
• Add the directory containing the Python DLL to your PATH environment variable, for example C:\Users\<You>\AppData\Local\Programs\Python\Python310
• If this happens when you are distributing your program, consider using PyOxidizer to package it with your binary.

If the wrong DLL is linked it is possible that this happened because another program added itself and its own Python DLLs to PATH. Rearrange your PATH variables to give the correct DLL priority.

Note: Changes to PATH (or any other environment variable) are not visible to existing shells. Restart it for changes to take effect.

For advanced troubleshooting, Dependency Walker can be used to diagnose linking errors.

Migrating from older PyO3 versions

This guide can help you upgrade code through breaking changes from one PyO3 version to the next. For a detailed list of all changes, see the CHANGELOG.

from 0.20.* to 0.21

Enable the gil-refs feature

# Cargo.toml
[dependencies]
pyo3 = "0.20"

# Cargo.toml
[dependencies]
pyo3 = { version = "0.21", features = ["gil-refs"] }

PyTypeInfo and PyTryFrom have been adjusted

#![allow(deprecated)]
use pyo3::prelude::*;
use pyo3::types::{PyInt, PyList};
fn main() -> PyResult<()> {
Python::with_gil(|py| {
    let list = PyList::new(py, 0..5);
    let b = <PyInt as PyTryFrom>::try_from(list.get_item(0).unwrap())?;
    Ok(())
})
}

use pyo3::prelude::*;
use pyo3::types::{PyInt, PyList};
fn main() -> PyResult<()> {
Python::with_gil(|py| {
    // Note that PyList::new is deprecated for PyList::new_bound as part of the GIL Refs API removal,
    // see the section below on migration to Bound<T>.
    #[allow(deprecated)]
    let list = PyList::new(py, 0..5);
    let b = list.get_item(0).unwrap().downcast::<PyInt>()?;
    Ok(())
})
}

Iter(A)NextOutput are deprecated

#![allow(deprecated)]
use pyo3::prelude::*;
use pyo3::iter::IterNextOutput;

#[pyclass]
struct PyClassIter {
    count: usize,
}

#[pymethods]
impl PyClassIter {
    fn __next__(&mut self) -> IterNextOutput<usize, &'static str> {
        if self.count < 5 {
            self.count += 1;
            IterNextOutput::Yield(self.count)
        } else {
            IterNextOutput::Return("done")
        }
    }
}

use pyo3::prelude::*;

#[pyclass]
struct PyClassIter {
    count: usize,
}

#[pymethods]
impl PyClassIter {
    fn __next__(&mut self) -> Option<usize> {
        if self.count < 5 {
            self.count += 1;
            Some(self.count)
        } else {
            None
        }
    }
}

use pyo3::prelude::*;
use pyo3::exceptions::PyStopIteration;

#[pyclass]
struct PyClassIter {
    count: usize,
}

#[pymethods]
impl PyClassIter {
    fn __next__(&mut self) -> PyResult<usize> {
        if self.count < 5 {
            self.count += 1;
            Ok(self.count)
        } else {
            Err(PyStopIteration::new_err("done"))
        }
    }
}

use pyo3::prelude::*;

#[pyclass]
struct PyClassAwaitable {
    number: usize,
}

#[pymethods]
impl PyClassAwaitable {
    fn __next__(&self) -> usize {
        self.number
    }

    fn __await__(slf: Py<Self>) -> Py<Self> {
        slf
    }
}

#[pyclass]
struct PyClassAsyncIter {
    number: usize,
}

#[pymethods]
impl PyClassAsyncIter {
    fn __anext__(&mut self) -> PyClassAwaitable {
        self.number += 1;
        PyClassAwaitable {
            number: self.number,
        }
    }

    fn __aiter__(slf: Py<Self>) -> Py<Self> {
        slf
    }
}

PyType::name has been renamed to PyType::qualname

PyCell has been deprecated

Migrating from the GIL Refs API to Bound<T>

let gil_ref: &PyAny = ...;
let bound: &Bound<PyAny> = &gil_ref.as_borrowed();

let obj: Py<PyList> = ...;
let bound: &Bound<'py, PyList> = obj.bind(py);
let bound: Bound<'py, PyList> = obj.into_bound(py);

let obj: &Py<PyList> = bound.as_unbound();
let obj: Py<PyList> = bound.unbind();

impl<'py> FromPyObject<'py> for MyType {
    fn extract(obj: &'py PyAny) -> PyResult<Self> {
        /* ... */
    }
}

impl<'py> FromPyObject<'py> for MyType {
    fn extract_bound(obj: &Bound<'py, PyAny>) -> PyResult<Self> {
        /* ... */
    }
}

Deactivating the gil-refs feature

#[cfg(feature = "gil-refs-migration")] {
use pyo3::prelude::*;
use pyo3::types::{PyList, PyType};
fn example<'py>(py: Python<'py>) -> PyResult<()> {
#[allow(deprecated)] // GIL Ref API
let obj: &'py PyType = py.get_type::<PyList>();
let name: &'py str = obj.getattr("__name__")?.extract()?;
assert_eq!(name, "list");
Ok(())
}
Python::with_gil(example).unwrap();
}

#[cfg(any(not(Py_LIMITED_API), Py_3_10))] {
use pyo3::prelude::*;
use pyo3::types::{PyList, PyType};
fn example<'py>(py: Python<'py>) -> PyResult<()> {
let obj: Bound<'py, PyType> = py.get_type_bound::<PyList>();
let name_obj: Bound<'py, PyAny> = obj.getattr("__name__")?;
// the lifetime of the data is no longer `'py` but the much shorter
// lifetime of the `name_obj` smart pointer above
let name: &'_ str = name_obj.extract()?;
assert_eq!(name, "list");
Ok(())
}
Python::with_gil(example).unwrap();
}

use pyo3::prelude::*;
use pyo3::types::{PyList, PyType};
fn example<'py>(py: Python<'py>) -> PyResult<()> {
use pyo3::pybacked::PyBackedStr;
let obj: Bound<'py, PyType> = py.get_type_bound::<PyList>();
let name: PyBackedStr = obj.getattr("__name__")?.extract()?;
assert_eq!(&*name, "list");
Ok(())
}
Python::with_gil(example).unwrap();

use pyo3::prelude::*;
use pyo3::exceptions::PyTypeError;
use pyo3::types::{PyDict, IntoPyDict};

fn main() {
let _ =
Python::with_gil(|py| {
    let dict: &PyDict = [("a", 1)].into_py_dict(py);
    // `a` is in the dictionary, with value 1
    assert!(dict.get_item("a").map_or(Ok(false), |x| x.eq(1))?);
    // `b` is not in the dictionary
    assert!(dict.get_item("b").is_none());
    // `dict` is not hashable, so this fails with a `TypeError`
    assert!(dict
        .get_item_with_error(dict)
        .unwrap_err()
        .is_instance_of::<PyTypeError>(py));
});
}

use pyo3::prelude::*;
use pyo3::exceptions::PyTypeError;
use pyo3::types::{PyDict, IntoPyDict};

fn main() {
let _ =
Python::with_gil(|py| -> PyResult<()> {
    let dict: &PyDict = [("a", 1)].into_py_dict(py);
    // `a` is in the dictionary, with value 1
    assert!(dict.get_item("a")?.map_or(Ok(false), |x| x.eq(1))?);
    // `b` is not in the dictionary
    assert!(dict.get_item("b")?.is_none());
    // `dict` is not hashable, so this fails with a `TypeError`
    assert!(dict
        .get_item(dict)
        .unwrap_err()
        .is_instance_of::<PyTypeError>(py));

    Ok(())
});
}

#[pyfunction]
fn x_or_y(x: Option<u64>, y: u64) -> u64 {
    x.unwrap_or(y)
}

#![allow(dead_code)]
use pyo3::prelude::*;

#[pyfunction]
#[pyo3(signature = (x, y))] // both x and y have no defaults and are required
fn x_or_y(x: Option<u64>, y: u64) -> u64 {
    x.unwrap_or(y)
}

#[pyfunction]
#[pyo3(a, b = "0", "/")]
fn add(a: u64, b: u64) -> u64 {
    a + b
}

#![allow(dead_code)]
use pyo3::prelude::*;

#[pyfunction]
#[pyo3(signature = (a, b=0, /))]
fn add(a: u64, b: u64) -> u64 {
    a + b
}

use pyo3::prelude::*;

#[pyclass]
struct SomeClass {}

impl SomeClass {
    fn __traverse__(&self, pyo3::class::gc::PyVisit<'_>) -> Result<(), pyo3::class::gc::PyTraverseError>` {
        Python::with_gil(|| { /*...*/ })  // ERROR: this will panic
    }
}

#[cfg(feature = "anyhow")]
#[allow(dead_code)]
mod anyhow_only {
use pyo3::prelude::*;
use pyo3::exceptions::PyValueError;
#[pyfunction]
fn raise_err() -> anyhow::Result<()> {
    Err(PyValueError::new_err("original error message").into())
}

fn main() {
    Python::with_gil(|py| {
        let rs_func = wrap_pyfunction!(raise_err, py).unwrap();
        pyo3::py_run!(
            py,
            rs_func,
            r"
        try:
            rs_func()
        except Exception as e:
            print(repr(e))
        "
        );
    })
}
}

#![allow(dead_code, deprecated)]
use pyo3::prelude::*;

#[pyclass]
struct SomeClass {}

struct ObjectAndGuard {
    object: Py<SomeClass>,
    guard: GILGuard,
}

impl ObjectAndGuard {
    fn new() -> Self {
        let guard = Python::acquire_gil();
        let object = Py::new(guard.python(), SomeClass {}).unwrap();

        Self { object, guard }
    }
}

let first = ObjectAndGuard::new();
let second = ObjectAndGuard::new();
// Panics because the guard within `second` is still alive.
drop(first);
drop(second);

#![allow(dead_code)]
use pyo3::prelude::*;

#[pyclass]
struct SomeClass {}

struct Object {
    object: Py<SomeClass>,
}

impl Object {
    fn new(py: Python<'_>) -> Self {
        let object = Py::new(py, SomeClass {}).unwrap();

        Self { object }
    }
}

// It either forces us to release the GIL before aquiring it again.
let first = Python::with_gil(|py| Object::new(py));
let second = Python::with_gil(|py| Object::new(py));
drop(first);
drop(second);

// Or it ensure releasing the inner lock before the outer one.
Python::with_gil(|py| {
    let first = Object::new(py);
    let second = Python::with_gil(|py| Object::new(py));
    drop(first);
    drop(second);
});

#![allow(dead_code)]
use pyo3::prelude::*;

#[pyfunction]
fn required_argument_after_option(x: Option<i32>, y: i32) {}

#![allow(dead_code)]
use pyo3::prelude::*;

// If x really was intended to be required
#[pyfunction(signature = (x, y))]
fn required_argument_after_option_a(x: Option<i32>, y: i32) {}

// If x was intended to be optional, y needs a default too
#[pyfunction(signature = (x=None, y=0))]
fn required_argument_after_option_b(x: Option<i32>, y: i32) {}

use pyo3::prelude::*;

// The `text_signature` option here is no longer necessary, as PyO3 will automatically
// generate exactly the same value.
#[pyfunction(text_signature = "(a, b, c)")]
fn simple_function(a: i32, b: i32, c: i32) {}

// The `text_signature` still provides value here as of PyO3 0.18, because the automatically
// generated signature would be "(a, b=..., c=...)".
#[pyfunction(signature = (a, b = 1, c = 2), text_signature = "(a, b=1, c=2)")]
fn function_with_defaults(a: i32, b: i32, c: i32) {}

fn main() {
    Python::with_gil(|py| {
        let simple = wrap_pyfunction_bound!(simple_function, py).unwrap();
        assert_eq!(simple.getattr("__text_signature__").unwrap().to_string(), "(a, b, c)");
        let defaulted = wrap_pyfunction_bound!(function_with_defaults, py).unwrap();
        assert_eq!(defaulted.getattr("__text_signature__").unwrap().to_string(), "(a, b=1, c=2)");
    })
}

use pyo3::prelude::*;
use std::collections::HashMap;

#[pyclass(mapping)]
struct Mapping {
    index: HashMap<String, usize>,
}

#[pymethods]
impl Mapping {
    #[new]
    fn new(elements: Option<&PyList>) -> PyResult<Self> {
    // ...
    // truncated implementation of this mapping pyclass - basically a wrapper around a HashMap
}

let m = Py::new(py, Mapping { index }).unwrap();
assert!(m.as_ref(py).downcast::<PyMapping>().is_err());
PyMapping::register::<Mapping>(py).unwrap();
assert!(m.as_ref(py).downcast::<PyMapping>().is_ok());

use pyo3::prelude::*;

fn main() {
Python::with_gil(|py| {
    // Cannot infer either `Py<PyAny>` or `Py<PyString>`
    let _test = "test".into_py(py);
});
}

use pyo3::prelude::*;

fn main() {
Python::with_gil(|py| {
    let _test: Py<PyAny> = "test".into_py(py);
});
}

use pyo3::Python;
use pyo3::type_object::PyTypeObject;
use pyo3::types::PyType;

fn get_type_object<T: PyTypeObject>(py: Python<'_>) -> &PyType {
    T::type_object(py)
}

use pyo3::{Python, PyTypeInfo};
use pyo3::types::PyType;

fn get_type_object<T: PyTypeInfo>(py: Python<'_>) -> &PyType {
    T::type_object(py)
}

Python::with_gil(|py| { get_type_object::<pyo3::types::PyList>(py); });

use pyo3::prelude::*;
use pyo3::class::{PyObjectProtocol, PyIterProtocol};
use pyo3::types::PyString;

#[pyclass]
struct MyClass {}

#[pyproto]
impl PyObjectProtocol for MyClass {
    fn __str__(&self) -> &'static [u8] {
        b"hello, world"
    }
}

#[pyproto]
impl PyIterProtocol for MyClass {
    fn __iter__(slf: PyRef<self>) -> PyResult<&PyAny> {
        PyString::new(slf.py(), "hello, world").iter()
    }
}

use pyo3::prelude::*;
use pyo3::types::PyString;

#[pyclass]
struct MyClass {}

#[pymethods]
impl MyClass {
    fn __str__(&self) -> &'static [u8] {
        b"hello, world"
    }

    fn __iter__(slf: PyRef<self>) -> PyResult<&PyAny> {
        PyString::new(slf.py(), "hello, world").iter()
    }
}

class ExampleContainer:

    def __len__(self):
        return 5

    def __getitem__(self, idx: int) -> int:
        if idx < 0 or idx > 5:
            raise IndexError()
        return idx

mod foo {
    use pyo3::prelude::*;

    #[pymodule]
    fn private_submodule(_py: Python<'_>, m: &PyModule) -> PyResult<()> {
        Ok(())
    }
}

use pyo3::prelude::*;
use foo::*;

#[pymodule]
fn my_module(_py: Python<'_>, m: &PyModule) -> PyResult<()> {
    m.add_wrapped(wrap_pymodule!(private_submodule))?;
    Ok(())
}

mod foo {
    use pyo3::prelude::*;

    #[pymodule]
    pub(crate) fn private_submodule(_py: Python<'_>, m: &PyModule) -> PyResult<()> {
        Ok(())
    }
}

use pyo3::prelude::*;
use pyo3::wrap_pymodule;
use foo::*;

#[pymodule]
fn my_module(_py: Python<'_>, m: &PyModule) -> PyResult<()> {
    m.add_wrapped(wrap_pymodule!(private_submodule))?;
    Ok(())
}

#![allow(deprecated)]
use pyo3::{Python, types::PyList};

Python::with_gil(|py| {
    let list = PyList::new(py, &[1, 2, 3]);
    assert_eq!(list[0..2].to_string(), "[1, 2]");
});

use pyo3::prelude::*;
use pyo3::class::basic::PyObjectProtocol;

#[pyclass]
struct MyClass {}

#[pyproto]
impl PyObjectProtocol for MyClass {
    fn __bytes__(&self) -> &'static [u8] {
        b"hello, world"
    }
}

use pyo3::prelude::*;

#[pyclass]
struct MyClass {}

#[pymethods]
impl MyClass {
    fn __bytes__(&self) -> &'static [u8] {
        b"hello, world"
    }
}

let result: PyResult<()> = PyErr::new::<TypeError, _>("error message").into();

use pyo3::{PyResult, exceptions::PyTypeError};
let result: PyResult<()> = Err(PyTypeError::new_err("error message"));

let err: PyErr = TypeError::py_err("error message");

use pyo3::{PyErr, PyResult, Python, type_object::PyTypeObject};
use pyo3::exceptions::{PyBaseException, PyTypeError};
Python::with_gil(|py| -> PyResult<()> {
let err: PyErr = PyTypeError::new_err("error message");

// Uses Display for PyErr, new for PyO3 0.12
assert_eq!(err.to_string(), "TypeError: error message");

// Now possible to interact with exception instances, new for PyO3 0.12
let instance: &PyBaseException = err.instance(py);
assert_eq!(
    instance.getattr("__class__")?,
    PyTypeError::type_object(py).as_ref()
);
Ok(())
}).unwrap();

use pyo3::prelude::*;
struct MyPyObjectWrapper(PyObject);

impl FromPy<MyPyObjectWrapper> for PyObject {
    fn from_py(other: MyPyObjectWrapper, _py: Python<'_>) -> Self {
        other.0
    }
}

use pyo3::prelude::*;
#[allow(dead_code)]
struct MyPyObjectWrapper(PyObject);

impl IntoPy<PyObject> for MyPyObjectWrapper {
    fn into_py(self, _py: Python<'_>) -> PyObject {
        self.0
    }
}

use pyo3::prelude::*;
Python::with_gil(|py| {
let obj = PyObject::from_py(1.234, py);
})

use pyo3::prelude::*;
Python::with_gil(|py| {
let obj: PyObject = 1.234.into_py(py);
})

use pyo3::{AsPyRef, Py, types::PyList};
pyo3::Python::with_gil(|py| {
let list_py: Py<PyList> = PyList::empty(py).into();
let list_ref: &PyList = list_py.as_ref(py);
})

use pyo3::{Py, types::PyList};
pyo3::Python::with_gil(|py| {
let list_py: Py<PyList> = PyList::empty(py).into();
let list_ref: &PyList = list_py.as_ref(py);
})

pyo3::Python::with_gil(|py| {
py.None().get_refcnt();
})

pyo3::Python::with_gil(|py| {
py.None().get_refcnt(py);
})

use pyo3::ObjectProtocol;

pyo3::Python::with_gil(|py| {
let obj = py.eval("lambda: 'Hi :)'", None, None).unwrap();
let hi: &pyo3::types::PyString = obj.call0().unwrap().downcast().unwrap();
assert_eq!(hi.len().unwrap(), 5);
})

pyo3::Python::with_gil(|py| {
let obj = py.eval("lambda: 'Hi :)'", None, None).unwrap();
let hi: &pyo3::types::PyString = obj.call0().unwrap().downcast().unwrap();
assert_eq!(hi.len().unwrap(), 5);
})

#[pyclass]
struct MyClass {}

#[pymethods]
impl MyClass {
    #[new]
    fn new(obj: &PyRawObject) {
        obj.init(MyClass {})
    }
}

use pyo3::prelude::*;
#[pyclass]
struct MyClass {}

#[pymethods]
impl MyClass {
    #[new]
    fn new() -> Self {
        MyClass {}
    }
}

use pyo3::prelude::*;

#[pyclass]
struct Names {
    names: Vec<String>,
}

#[pymethods]
impl Names {
    #[new]
    fn new() -> Self {
        Names { names: vec![] }
    }
    fn merge(&mut self, other: &mut Names) {
        self.names.append(&mut other.names)
    }
}
Python::with_gil(|py| {
    let names = Py::new(py, Names::new()).unwrap();
    pyo3::py_run!(py, names, r"
    try:
       names.merge(names)
       assert False, 'Unreachable'
    except RuntimeError as e:
       assert str(e) == 'Already borrowed'
    ");
})

use pyo3::prelude::*;
#[pyclass]
struct MyClass {}
Python::with_gil(|py| {
let obj_ref = PyRef::new(py, MyClass {}).unwrap();
})

use pyo3::prelude::*;
#[pyclass]
struct MyClass {}
Python::with_gil(|py| {
let obj = PyCell::new(py, MyClass {}).unwrap();
let obj_ref = obj.borrow();
})

let obj: &PyAny = create_obj();
let obj_ref: &MyClass = obj.extract().unwrap();
let obj_ref_mut: &mut MyClass = obj.extract().unwrap();

use pyo3::prelude::*;
use pyo3::types::IntoPyDict;
#[pyclass] #[derive(Clone)] struct MyClass {}
#[pymethods] impl MyClass { #[new]fn new() -> Self { MyClass {} }}
Python::with_gil(|py| {
let typeobj = py.get_type::<MyClass>();
let d = [("c", typeobj)].into_py_dict(py);
let create_obj = || py.eval("c()", None, Some(d)).unwrap();
let obj: &PyAny = create_obj();
let obj_cell: &PyCell<MyClass> = obj.extract().unwrap();
let obj_cloned: MyClass = obj.extract().unwrap(); // extracted by cloning the object
{
    let obj_ref: PyRef<'_, MyClass> = obj.extract().unwrap();
    // we need to drop obj_ref before we can extract a PyRefMut due to Rust's rules of references
}
let obj_ref_mut: PyRefMut<'_, MyClass> = obj.extract().unwrap();
})

use pyo3::prelude::*;
use pyo3::class::PySequenceProtocol;
#[pyclass]
struct ByteSequence {
    elements: Vec<u8>,
}
#[pyproto]
impl PySequenceProtocol for ByteSequence {
    fn __concat__(&self, other: &Self) -> PyResult<Self> {
        let mut elements = self.elements.clone();
        elements.extend_from_slice(&other.elements);
        Ok(Self { elements })
    }
}

use pyo3::prelude::*;
use pyo3::class::PySequenceProtocol;
#[pyclass]
struct ByteSequence {
    elements: Vec<u8>,
}
#[pyproto]
impl PySequenceProtocol for ByteSequence {
    fn __concat__(&self, other: PyRef<'p, Self>) -> PyResult<Self> {
        let mut elements = self.elements.clone();
        elements.extend_from_slice(&other.elements);
        Ok(Self { elements })
    }
}

#![allow(dead_code)]
pub trait Model {
    fn set_variables(&mut self, inputs: &Vec<f64>);
    fn compute(&mut self);
    fn get_results(&self) -> Vec<f64>;
}

pub fn solve<T: Model>(model: &mut T) {
    println!("Magic solver that mutates the model into a resolved state");
}

class Model:
    def set_variables(self, inputs):
        self.inputs = inputs
    def compute(self):
        self.results = [elt**2 - 3 for elt in self.inputs]
    def get_results(self):
        return self.results

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyList;

pub trait Model {
  fn set_variables(&mut self, inputs: &Vec<f64>);
  fn compute(&mut self);
  fn get_results(&self) -> Vec<f64>;
}

struct UserModel {
    model: Py<PyAny>,
}

impl Model for UserModel {
    fn set_variables(&mut self, var: &Vec<f64>) {
        println!("Rust calling Python to set the variables");
        Python::with_gil(|py| {
            self.model
                .bind(py)
                .call_method("set_variables", (PyList::new_bound(py, var),), None)
                .unwrap();
        })
    }

    fn get_results(&self) -> Vec<f64> {
        println!("Rust calling Python to get the results");
        Python::with_gil(|py| {
            self.model
                .bind(py)
                .call_method("get_results", (), None)
                .unwrap()
                .extract()
                .unwrap()
        })
    }

    fn compute(&mut self) {
        println!("Rust calling Python to perform the computation");
        Python::with_gil(|py| {
            self.model
                .bind(py)
                .call_method("compute", (), None)
                .unwrap();
        })
    }
}

#![allow(dead_code)]
pub trait Model {
  fn set_variables(&mut self, inputs: &Vec<f64>);
  fn compute(&mut self);
  fn get_results(&self) -> Vec<f64>;
}
use pyo3::prelude::*;

#[pyclass]
struct UserModel {
    model: Py<PyAny>,
}

#[pymodule]
fn trait_exposure(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_class::<UserModel>()?;
    Ok(())
}

#[pymethods]
impl UserModel {
    #[new]
    pub fn new(model: Py<PyAny>) -> Self {
        UserModel { model }
    }
}

#[pymethods]
impl Model for UserModel {
    // the previous trait implementation
}

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyList;

pub trait Model {
  fn set_variables(&mut self, inputs: &Vec<f64>);
  fn compute(&mut self);
  fn get_results(&self) -> Vec<f64>;
}

#[pyclass]
struct UserModel {
    model: Py<PyAny>,
}

impl Model for UserModel {
 fn set_variables(&mut self, var: &Vec<f64>) {
     println!("Rust calling Python to set the variables");
     Python::with_gil(|py| {
         self.model.bind(py)
             .call_method("set_variables", (PyList::new_bound(py, var),), None)
             .unwrap();
     })
 }

 fn get_results(&self) -> Vec<f64> {
     println!("Rust calling Python to get the results");
     Python::with_gil(|py| {
         self.model
             .bind(py)
             .call_method("get_results", (), None)
             .unwrap()
             .extract()
             .unwrap()
     })
 }

 fn compute(&mut self) {
     println!("Rust calling Python to perform the computation");
     Python::with_gil(|py| {
         self.model
             .bind(py)
             .call_method("compute", (), None)
             .unwrap();
     })

 }
}

#[pymethods]
impl UserModel {
    pub fn set_variables(&mut self, var: Vec<f64>) {
        println!("Set variables from Python calling Rust");
        Model::set_variables(self, &var)
    }

    pub fn get_results(&mut self) -> Vec<f64> {
        println!("Get results from Python calling Rust");
        Model::get_results(self)
    }

    pub fn compute(&mut self) {
        println!("Compute from Python calling Rust");
        Model::compute(self)
    }
}

class Model:
    def set_variables(self, inputs):
        self.inputs = inputs
    def compute(self):
        self.results = [elt**2 - 3 for elt in self.inputs]
    def get_results(self):
        return self.results

if __name__=="__main__":
  import trait_exposure

  myModel = Model()
  my_rust_model = trait_exposure.UserModel(myModel)
  my_rust_model.set_variables([2.0])
  print("Print value from Python: ", myModel.inputs)
  my_rust_model.compute()
  print("Print value from Python through Rust: ", my_rust_model.get_results())
  print("Print value directly from Python: ", myModel.get_results())

Set variables from Python calling Rust
Set variables from Rust calling Python
Print value from Python:  [2.0]
Compute from Python calling Rust
Compute from Rust calling Python
Get results from Python calling Rust
Get results from Rust calling Python
Print value from Python through Rust:  [1.0]
Print value directly from Python:  [1.0]

File "main.py", line 15, in <module>
   my_rust_model.set_variables(2)
TypeError

class Model:
    def set_variables(self, inputs):
        self.inputs = inputs
    def compute(self):
        self.results = [elt**2 -3 for elt in self.inputs]
    def get_results(self):
        return self.results[0]
        #return self.results <-- this is the expected output

pyo3_runtime.PanicException: called `Result::unwrap()` on an `Err` value: PyErr { type: Py(0x10dcf79f0, PhantomData) }

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyList;

pub trait Model {
  fn set_variables(&mut self, inputs: &Vec<f64>);
  fn compute(&mut self);
  fn get_results(&self) -> Vec<f64>;
}

#[pyclass]
struct UserModel {
    model: Py<PyAny>,
}

impl Model for UserModel {
    fn get_results(&self) -> Vec<f64> {
        println!("Rust calling Python to get the results");
        Python::with_gil(|py| {
            self.model
                .bind(py)
                .call_method("get_results", (), None)
                .unwrap()
                .extract()
                .unwrap()
        })
    }
    fn set_variables(&mut self, var: &Vec<f64>) {
        println!("Rust calling Python to set the variables");
        Python::with_gil(|py| {
            self.model.bind(py)
                .call_method("set_variables", (PyList::new_bound(py, var),), None)
                .unwrap();
        })
    }

    fn compute(&mut self) {
        println!("Rust calling Python to perform the computation");
        Python::with_gil(|py| {
            self.model
                .bind(py)
                .call_method("compute", (), None)
                .unwrap();
        })
    }
}

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyList;

pub trait Model {
  fn set_variables(&mut self, inputs: &Vec<f64>);
  fn compute(&mut self);
  fn get_results(&self) -> Vec<f64>;
}

#[pyclass]
struct UserModel {
    model: Py<PyAny>,
}

impl Model for UserModel {
    fn get_results(&self) -> Vec<f64> {
        println!("Get results from Rust calling Python");
        Python::with_gil(|py| {
            let py_result: Bound<'_, PyAny> = self
                .model
                .bind(py)
                .call_method("get_results", (), None)
                .unwrap();

            if py_result.get_type().name().unwrap() != "list" {
                panic!(
                    "Expected a list for the get_results() method signature, got {}",
                    py_result.get_type().name().unwrap()
                );
            }
            py_result.extract()
        })
        .unwrap()
    }
    fn set_variables(&mut self, var: &Vec<f64>) {
        println!("Rust calling Python to set the variables");
        Python::with_gil(|py| {
            let py_model = self.model.bind(py)
                .call_method("set_variables", (PyList::new_bound(py, var),), None)
                .unwrap();
        })
    }

    fn compute(&mut self) {
        println!("Rust calling Python to perform the computation");
        Python::with_gil(|py| {
            self.model
                .bind(py)
                .call_method("compute", (), None)
                .unwrap();
        })
    }
}

#![allow(dead_code)]
use pyo3::prelude::*;
use pyo3::types::PyList;

pub trait Model {
    fn set_variables(&mut self, var: &Vec<f64>);
    fn get_results(&self) -> Vec<f64>;
    fn compute(&mut self);
}

pub fn solve<T: Model>(model: &mut T) {
    println!("Magic solver that mutates the model into a resolved state");
}

#[pyfunction]
#[pyo3(name = "solve")]
pub fn solve_wrapper(model: &mut UserModel) {
    solve(model);
}

#[pyclass]
pub struct UserModel {
    model: Py<PyAny>,
}

#[pymodule]
fn trait_exposure(m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_class::<UserModel>()?;
    m.add_function(wrap_pyfunction!(solve_wrapper, m)?)?;
    Ok(())
}

#[pymethods]
impl UserModel {
    #[new]
    pub fn new(model: Py<PyAny>) -> Self {
        UserModel { model }
    }

    pub fn set_variables(&mut self, var: Vec<f64>) {
        println!("Set variables from Python calling Rust");
        Model::set_variables(self, &var)
    }

    pub fn get_results(&mut self) -> Vec<f64> {
        println!("Get results from Python calling Rust");
        Model::get_results(self)
    }

    pub fn compute(&mut self) {
        Model::compute(self)
    }
}

impl Model for UserModel {
    fn set_variables(&mut self, var: &Vec<f64>) {
        println!("Rust calling Python to set the variables");
        Python::with_gil(|py| {
            self.model
                .bind(py)
                .call_method("set_variables", (PyList::new_bound(py, var),), None)
                .unwrap();
        })
    }

    fn get_results(&self) -> Vec<f64> {
        println!("Get results from Rust calling Python");
        Python::with_gil(|py| {
            let py_result: Bound<'_, PyAny> = self
                .model
                .bind(py)
                .call_method("get_results", (), None)
                .unwrap();

            if py_result.get_type().name().unwrap() != "list" {
                panic!(
                    "Expected a list for the get_results() method signature, got {}",
                    py_result.get_type().name().unwrap()
                );
            }
            py_result.extract()
        })
        .unwrap()
    }

    fn compute(&mut self) {
        println!("Rust calling Python to perform the computation");
        Python::with_gil(|py| {
            self.model
                .bind(py)
                .call_method("compute", (), None)
                .unwrap();
        })
    }
}