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.

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 and PyPy)
• Rust 1.48 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 = "2018"

[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 = "0.18.3", features = ["extension-module"] }

src/lib.rs

#![allow(unused)]
fn main() {
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(_py: Python<'_>, m: &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 = "0.18.3"
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("sys")?;
        let version: String = sys.getattr("version")?.extract()?;

        let locals = [("os", py.import("os")?)].into_py_dict(py);
        let code = "os.getenv('USER') or os.getenv('USERNAME') or 'Unknown'";
        let user: String = py.eval(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.
• 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.
• jsonschema-rs Fast JSON Schema validation library.
• mocpy Astronomical Python library offering data structures for describing any arbitrary coverage regions on the unit sphere.
• 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.
• wasmer-python Python library to run WebAssembly binaries.

Articles and other media

• 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 Gitter
• 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. Python is licensed under the Python License.

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.

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 1.48.

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 that when using pyenv, you should also set the following environment variable:

PYTHON_CONFIGURE_OPTS="--enable-shared"

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. maturin is just a Python package, so you can add it in the same 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 -D 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 = "0.18.3", features = ["extension-module"] }

pyproject.toml

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

[build-system]
requires = ["maturin>=0.14,<0.15"]
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:

#![allow(unused)]
fn main() {
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(_py: Python<'_>, m: &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.

Python modules

You can create a module using #[pymodule]:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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

/// This module is implemented in Rust.
#[pymodule]
fn my_extension(py: Python<'_>, m: &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")]:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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

#[pymodule]
#[pyo3(name = "custom_name")]
fn my_extension(py: Python<'_>, m: &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 PyModule.add_submodule(). For example, you could define the modules parent_module and parent_module.child_module.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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

fn register_child_module(py: Python<'_>, parent_module: &PyModule) -> PyResult<()> {
    let child_module = PyModule::new(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(py);

   py.run("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.

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:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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

#[pymodule]
fn my_extension(py: Python<'_>, m: &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:

  #![allow(unused)]
  fn main() {
  use pyo3::prelude::*;
  
  #[pyfunction]
  #[pyo3(name = "no_args")]
  fn no_args_py() -> usize {
      42
  }
  
  #[pymodule]
  fn module_with_functions(py: Python<'_>, m: &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 &PyModule.

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

  #![allow(unused)]
  fn main() {
  use pyo3::prelude::*;
  
  #[pyfunction]
  #[pyo3(pass_module)]
  fn pyfunction_with_module(module: &PyModule) -> PyResult<&str> {
      module.name()
  }
  
  #[pymodule]
  fn module_with_fn(py: Python<'_>, m: &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(&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:

  #![allow(unused)]
  fn main() {
  use pyo3::prelude::*;
  
  fn get_length(obj: &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!(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 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 PyAny::call with the args as first parameter and the kwargs (or None) as second parameter. There are also PyAny::call0 with no args and 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 PyCFunction given a #[pyfunction] and a 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:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

#[pymodule]
fn my_extension(py: Python<'_>, m: &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:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

#[pymodule]
fn my_extension(py: Python<'_>, m: &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. There are two ways to modify this behaviour:

• The #[pyo3(signature = (...))] option which allows writing a signature in Python syntax.
• Extra arguments directly to #[pyfunction]. (See deprecated form)

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:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
use pyo3::types::PyDict;

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

#[pymodule]
fn module_with_functions(py: Python<'_>, m: &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 &PyTuple.
• **kwargs: "kwargs" receives keyword arguments. The type of the kwargs parameter has to be Option<&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:

#![allow(unused)]
fn main() {
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: &PyTuple,
        name: &str,
        py_kwargs: Option<&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(unused)]
fn main() {
#![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(unused)]
fn main() {
#![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!(increment, py)?;

        let inspect = PyModule::import(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=Ellipsis)");

        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!(increment, py)?;

        let inspect = PyModule::import(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>.

Deprecated form

The #[pyfunction] macro can take the argument specification directly, but this method is deprecated in PyO3 0.18 because the #[pyo3(signature)] option offers a simpler syntax and better validation.

The #[pymethods] macro has an #[args] attribute which accepts the deprecated form.

Below are the same examples as above which using the deprecated syntax:

#![allow(unused)]
fn main() {
#![allow(deprecated)]

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

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

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

The following parameters can be passed to the #[pyfunction] attribute:

• "/": positional-only arguments separator, each parameter defined before "/" is a positional-only parameter. Corresponds to python's def meth(arg1, arg2, ..., /, argN..).
• "*": var arguments separator, each parameter defined after "*" is a keyword-only parameter. Corresponds to python's def meth(*, arg1.., arg2=..).
• args="*": "args" is var args, corresponds to Python's def meth(*args). Type of the args parameter has to be &PyTuple.
• kwargs="**": "kwargs" receives keyword arguments, corresponds to Python's def meth(**kwargs). The type of the kwargs parameter has to be Option<&PyDict>.
• arg="Value": arguments with default value. Corresponds to Python's def meth(arg=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:

#![allow(unused)]
fn main() {
#![allow(deprecated)]
use pyo3::prelude::*;
use pyo3::types::{PyDict, PyTuple};

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

    #[args(num = "10", py_args = "*", name = "\"Hello\"", py_kwargs = "**")]
    fn method(
        &mut self,
        num: i32,
        py_args: &PyTuple,
        name: &str,
        py_kwargs: Option<&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))
    }
}
}

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 has some limitations, which may be improved in the future:

• It will not include the value of default arguments, replacing them all with .... (.pyi type stub files commonly also use ... for all default arguments in the same way.)
• Nothing is generated for the #[new] method of a #[pyclass].

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!(add, py)?;

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

        let inspect = PyModule::import(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=Ellipsis, /)");

        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=Ellipsis, /)
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!(add, py)?;

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

        let inspect = PyModule::import(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!(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 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 implementing a conversion 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!(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!(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!(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!(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 a fieldless enum (a.k.a. C-like 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

Defining a new class

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

#![allow(unused)]
fn main() {
#![allow(dead_code)]
use pyo3::prelude::*;

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

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

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

#[pyclass]
enum MyEnum {
    Variant,
    OtherVariant = 30, // PyO3 supports custom discriminants.
}
}

The above example generates implementations for PyTypeInfo and PyClass for MyClass and MyEnum. 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 which 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.

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(unused)]
fn main() {
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(unused)]
fn main() {
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))
        }
    }
}
}

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(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
struct Number(i32);

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

PyCell and interior mutability

You sometimes need to convert your pyclass into a Python object and access it from Rust code (e.g., for testing it). PyCell is the primary interface for that.

PyCell<T: PyClass> is always allocated in the Python heap, so Rust doesn't have ownership of it. In other words, Rust code can only extract a &PyCell<T>, not a PyCell<T>.

Thus, to mutate data behind &PyCell safely, PyO3 employs the Interior Mutability Pattern like RefCell.

Users who are familiar with RefCell can use PyCell just like RefCell.

For users who are not very familiar with RefCell, 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 must always be valid.

PyCell, like RefCell, ensures these borrowing rules by tracking references at runtime.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
struct MyClass {
    #[pyo3(get)]
    num: i32,
}
Python::with_gil(|py| {
    let obj = PyCell::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 `&PyCell` to a Python object
    pyo3::py_run!(py, obj, "assert obj.num == 5");
});
}

&PyCell<T> is bounded by the same lifetime as a GILGuard. To make the object longer lived (for example, to store it in a struct on the Rust side), you can use Py<T>, which stores an object longer than the GIL lifetime, and therefore needs a Python<'_> token to access.

#![allow(unused)]
fn main() {
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 cell = obj.as_ref(py); // Py<MyClass>::as_ref returns &PyCell<MyClass>
    let obj_ref = cell.borrow(); // Get PyRef<T>
    assert_eq!(obj_ref.num, 1);
});
}

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.

For convenience, (T, U) implements Into<PyClassInitializer<T>> where U is the base class of T. But for 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>.

#![allow(unused)]
fn main() {
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)
    }
}
Python::with_gil(|py| {
    let subsub = pyo3::PyCell::new(py, SubSubClass::new()).unwrap();
    pyo3::py_run!(py, subsub, "assert subsub.method3() == 3000")
});
}

You can also 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).

However, because of some technical problems, we don't currently provide safe upcasting methods for types that inherit native types. Even in such cases, you can unsafely get a base class by raw pointer conversion.

#![allow(unused)]
fn main() {
#[cfg(not(Py_LIMITED_API))] {
use pyo3::prelude::*;
use pyo3::types::PyDict;
use pyo3::AsPyPointer;
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(mut self_: PyRefMut<'_, Self>, key: String, value: &PyAny) -> PyResult<()> {
        self_.counter.entry(key.clone()).or_insert(0);
        let py = self_.py();
        let dict: &PyDict = unsafe { py.from_borrowed_ptr_or_err(self_.as_ptr())? };
        dict.set_item(key, value)
    }
}
Python::with_gil(|py| {
    let cnt = pyo3::PyCell::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.

#![allow(unused)]
fn main() {
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(unused)]
fn main() {
#[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: &PyAny, kwargs: Option<&PyAny>) -> Self {
        Self { private: 0 }
    }

    // some custom methods that use `private` here...
}
Python::with_gil(|py| {
    let cls = py.get_type::<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:

#![allow(unused)]
fn main() {
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:

#![allow(unused)]
fn main() {
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).

#![allow(unused)]
fn main() {
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.

#![allow(unused)]
fn main() {
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.

#![allow(unused)]
fn main() {
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.

#![allow(unused)]
fn main() {
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.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
use pyo3::types::PyType;
#[pyclass]
struct MyClass {
    #[allow(dead_code)]
    num: i32,
}
#[pymethods]
impl MyClass {
    #[classmethod]
    fn cls_method(cls: &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 &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>.

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>.

#![allow(unused)]
fn main() {
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.

#![allow(unused)]
fn main() {
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::<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:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
struct MyClass {}
#[pymethods]
impl MyClass {
    #[classattr]
    const MY_CONST_ATTRIBUTE: &'static str = "foobar";
}
}

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 which 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:

#![allow(unused)]
fn main() {
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: &PyTuple,
        name: &str,
        py_kwargs: Option<&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

Making class method signatures available to Python

The text_signature = "..." option for #[pyfunction] also works for classes and methods:

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

// it works even if the item is not documented:
#[pyclass(text_signature = "(c, d, /)")]
struct MyClass {}

#[pymethods]
impl MyClass {
    // the signature for the constructor is attached
    // to the struct definition instead.
    #[new]
    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
    }
    #[classmethod]
    #[pyo3(text_signature = "(cls, e, f)")]
    fn my_class_method(cls: &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(py, "inspect")?.getattr("signature")?;
        let module = PyModule::new(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, "(cls, e, f)");
        }

        {
            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 classes is not compatible with compilation in abi3 mode until Python 3.10 or greater.

#[pyclass] enums

Currently PyO3 only supports fieldless enums. 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 ==:

#![allow(unused)]
fn main() {
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::<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 enums into int:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
enum MyEnum {
    Variant,
    OtherVariant = 10,
}

Python::with_gil(|py| {
    let cls = py.get_type::<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:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
enum MyEnum{
    Variant,
    OtherVariant,
}

Python::with_gil(|py| {
    let cls = py.get_type::<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__:

#![allow(unused)]
fn main() {
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::<MyEnum>();
    pyo3::py_run!(py, cls, "assert repr(cls.Answer) == '42'")
})
}

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

#![allow(unused)]
fn main() {
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::<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.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass(subclass)]
enum BadBase {
    Var1,
}
}

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

#[pyclass(subclass)]
struct Base;

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

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

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:

#![allow(unused)]
fn main() {
#[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,
}
unsafe impl pyo3::type_object::PyTypeInfo for MyClass {
    type AsRefTarget = pyo3::PyCell<Self>;
    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: &'py pyo3::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: &'py pyo3::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 DOC: &'static str = "Class for demonstration\u{0}";
    const IS_BASETYPE: bool = false;
    const IS_SUBCLASS: bool = false;
    type Layout = PyCell<MyClass>;
    type BaseType = PyAny;
    type ThreadChecker = pyo3::impl_::pyclass::ThreadCheckerStub<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
    }
}

Python::with_gil(|py| {
    let cls = py.get_type::<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 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 PyCell 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:

  #![allow(unused)]
  fn main() {
  use pyo3::prelude::*;
  
  #[pyclass]
  struct NotHashable {}
  
  #[pymethods]
  impl NotHashable {
      #[classattr]
      const __hash__: Option<PyObject> = None;
  }
  }
  

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

  Overloads Python comparison operations (==, !=, <, <=, >, and >=). The CompareOp argument indicates the comparison operation being performed.

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

  You can use CompareOp::matches to adapt a Rust std::cmp::Ordering result to the requested comparison.

• __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:

#![allow(unused)]
fn main() {
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:

#![allow(unused)]
fn main() {
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::PyCell::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:

  #![allow(unused)]
  fn main() {
  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:

#![allow(unused)]
fn main() {
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;
    }
}
}

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(unused)]
fn main() {
use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

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

#[pymodule]
fn my_module(_py: Python<'_>, m: &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.

#![allow(unused)]
fn main() {
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 PyCell as the self argument.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

#[pymethods]
impl Number {
    fn __repr__(slf: &PyCell<Self>) -> PyResult<String> {
        // This is the equivalent of `self.__class__.__name__` in Python.
        let class_name: &str = slf.get_type().name()?;
        // 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.

#![allow(unused)]
fn main() {
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:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
struct NotHashable {}

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

Comparisons

Unlike in Python, PyO3 does not provide the magic comparison methods you might expect like __eq__, __lt__ and so on. Instead you have to implement all six operations at once with __richcmp__. This method will be called with a value of CompareOp depending on the operation.

#![allow(unused)]
fn main() {
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:

#![allow(unused)]
fn main() {
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, if you want to leave some operations unimplemented, you can return py.NotImplemented() for some of the operations:

#![allow(unused)]
fn main() {
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(),
        }
    }
}
}

Truthyness

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

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

#[pyclass]
struct Number(i32);

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

Final code

#![allow(unused)]
fn main() {
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: &PyCell<Self>) -> PyResult<String> {
        let class_name: &str = slf.get_type().name()?;
        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(_py: Python<'_>, m: &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(unused)]
fn main() {
#![allow(dead_code)]
use pyo3::prelude::*;

fn wrap(obj: &PyAny) -> Result<i32, PyErr> {
    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 and the #[pyo3(text_signature = "...")] attribute, both of which are visible to Python users.

#![allow(unused)]
fn main() {
#![allow(dead_code)]
use pyo3::prelude::*;

fn wrap(obj: &PyAny) -> Result<i32, PyErr> {
    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")]
#[pyo3(text_signature = "(int)")]
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:

#![allow(unused)]
fn main() {
use std::convert::TryInto;
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

#![allow(unused)]
fn main() {
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.

#![allow(unused)]
fn main() {
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>) -> &'py PyComplex {
        PyComplex::from_doubles(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 std::convert::TryInto;

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

fn wrap(obj: &PyAny) -> Result<i32, PyErr> {
    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")]
#[pyo3(text_signature = "(int)")]
struct Number(i32);

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

    fn __repr__(slf: &PyCell<Self>) -> PyResult<String> {
       // Get the class name dynamically in case `Number` is subclassed
       let class_name: &str = slf.get_type().name()?;
        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>) -> &'py PyComplex {
        PyComplex::from_doubles(py, self.0 as f64, 0.0)
    }
}

#[pymodule]
fn my_module(_py: Python<'_>, m: &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__() == '(int)'
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(py, "__main__")?.dict();
        globals.set_item("Number", Number::type_object(py))?;

        py.run(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(&PyAny) -> PyResult<T>.

• &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(unused)]
fn main() {
#![allow(dead_code)]
use std::os::raw::c_ulong;
use pyo3::prelude::*;
use pyo3::ffi;
use pyo3::conversion::AsPyPointer;

fn wrap(obj: &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: &PyTuple,
        kwargs: Option<&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(py, args, kwargs)?;

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

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

Python code:

@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: &PyTuple, kwargs: Option<&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.

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:

There are also a few special types related to the GIL and Rust-defined #[pyclass]es which may come in useful:

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, Python-native types (&PyAny, &PyDict etc.) can be used with zero cost.

Because these types are references, in some situations the Rust compiler may ask for lifetime annotations. If this is the case, you should use Py<PyAny>, Py<PyDict> etc. instead - which are also zero-cost. For all of these Python-native types T, Py<T> can be created from T with an .into() conversion.

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(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(
            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(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(
            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.

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(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(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(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(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(
                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(
                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(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(&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.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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.

Python exceptions

Defining a new exception

You can use the create_exception! macro to define a new exception type:

#![allow(unused)]
fn main() {
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:

#![allow(unused)]
fn main() {
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::<CustomError>())].into_py_dict(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:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
use pyo3::exceptions::PyException;

pyo3::create_exception!(mymodule, CustomError, PyException);

#[pymodule]
fn mymodule(py: Python<'_>, m: &PyModule) -> PyResult<()> {
    // ... other elements added to module ...
    m.add("CustomError", py.get_type::<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:

#![allow(unused)]
fn main() {
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.

#![allow(unused)]
fn main() {
use pyo3::Python;
use pyo3::types::{PyBool, PyList};

Python::with_gil(|py| {
    assert!(PyBool::new(py, true).is_instance_of::<PyBool>().unwrap());
    let list = PyList::new(py, &[1, 2, 3, 4]);
    assert!(!list.is_instance_of::<PyBool>().unwrap());
    assert!(list.is_instance_of::<PyList>().unwrap());
});
}

To check the type of an exception, you can similarly do:

#![allow(unused)]
fn main() {
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(unused)]
#![allow(dead_code)]
fn main() {
use pyo3::prelude::*;

mod io {
    pyo3::import_exception!(io, UnsupportedOperation);
}

fn tell(file: &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 in Rust code

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

• How to call Python functions
• How to execute existing Python code

Calling Python functions

Any Python-native object reference (such as &PyAny, &PyList, or &PyCell<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(
            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)?;

        // call object with PyTuple
        let args = PyTuple::new(py, &[arg1, arg2, arg3]);
        fun.call1(py, args)?;

        // pass arguments as rust tuple
        let args = (arg1, arg2, arg3);
        fun.call1(py, args)?;
        Ok(())
    })
}

Creating keyword arguments

For the call and call_method APIs, kwargs can be None or Some(&PyDict). 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(
            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(py);
        fun.call(py, (), Some(kwargs))?;

        // pass arguments as Vec
        let kwargs = vec![(key1, val1), (key2, val2)];
        fun.call(py, (), Some(kwargs.into_py_dict(py)))?;

        // pass arguments as HashMap
        let mut kwargs = HashMap::<&str, i32>::new();
        kwargs.insert(key1, 1);
        fun.call(py, (), Some(kwargs.into_py_dict(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(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 &PyAny object.

use pyo3::prelude::*;

fn main() -> Result<(), ()> {
Python::with_gil(|py| {
    let result = py
        .eval("[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::{PyCell, 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 = PyCell::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, PyModule}};

fn main() -> PyResult<()> {
Python::with_gil(|py| {
    let activators = PyModule::from_code(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(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(_py: Python<'_>, foo_module: &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(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(py, "foo")?;
        foo_module.add_function(wrap_pyfunction!(add_one, foo_module)?)?;

        // Import and get sys.modules
        let sys = PyModule::import(py, "sys")?;
        let py_modules: &PyDict = sys.getattr("modules")?.downcast()?;

        // Insert foo into sys.modules
        py_modules.set_item("foo", foo_module)?;

        // Now we can import + run our python code
        Python::run(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(py, py_foo, "utils.foo", "utils.foo")?;
        let app: Py<PyAny> = PyModule::from_code(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: &PyList = py.import("sys")?.getattr("path")?.downcast()?;
        syspath.insert(0, &path)?;
        let app: Py<PyAny> = PyModule::from_code(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::*;
use pyo3::types::PyModule;

fn main() {
    Python::with_gil(|py| {
        let custom_manager = PyModule::from_code(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("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(py), e.value(py), e.traceback(py))
                ).unwrap();
            }
        }
    })
}

GIL lifetimes, mutability and Python object types

On first glance, PyO3 provides a huge number of different types that can be used to wrap or refer to Python objects. This page delves into the details and gives an overview of their intended meaning, with examples when each type is best used.

The Python GIL, mutability, and Rust types

Since Python has no concept of ownership, and works solely with boxed objects, any Python object can be referenced any number of times, and mutation is allowed from any reference.

The situation is helped a little by the Global Interpreter Lock (GIL), which ensures that only one thread can use the Python interpreter and its API at the same time, while non-Python operations (system calls and extension code) can unlock the GIL. (See the section on parallelism for how to do that in PyO3.)

In PyO3, holding the GIL is modeled by acquiring a token of the type Python<'py>, which serves three purposes:

• It provides some global API for the Python interpreter, such as eval.
• It can be passed to functions that require a proof of holding the GIL, such as Py::clone_ref.
• Its lifetime can be used to create Rust references that implicitly guarantee holding the GIL, such as &'py PyAny.

The latter two points are the reason why some APIs in PyO3 require the py: Python argument, while others don't.

The PyO3 API for Python objects is written such that instead of requiring a mutable Rust reference for mutating operations such as PyList::append, a shared reference (which, in turn, can only be created through Python<'_> with a GIL lifetime) is sufficient.

However, Rust structs wrapped as Python objects (called pyclass types) usually do need &mut access. Due to the GIL, PyO3 can guarantee thread-safe access to them, but it cannot statically guarantee uniqueness of &mut references once an object's ownership has been passed to the Python interpreter, ensuring references is done at runtime using PyCell, a scheme very similar to std::cell::RefCell.

Accessing the Python GIL

To get hold of a Python<'py> token to prove the GIL is held, consult PyO3's documentation.

Object types

PyAny

Represents: a Python object of unspecified type, restricted to a GIL lifetime. Currently, PyAny can only ever occur as a reference, &PyAny.

Used: Whenever you want to refer to some Python object and will have the GIL for the whole duration you need to access that object. For example, intermediate values and arguments to pyfunctions or pymethods implemented in Rust where any type is allowed.

Many general methods for interacting with Python objects are on the PyAny struct, such as getattr, setattr, and .call.

Conversions:

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

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
use pyo3::types::PyList;
Python::with_gil(|py| -> PyResult<()> {
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]:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass] #[derive(Clone)] struct MyClass { }
Python::with_gil(|py| -> PyResult<()> {
let obj: &PyAny = Py::new(py, MyClass {})?.into_ref(py);

// To &PyCell<MyClass> with PyAny::downcast
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

Represents: a native Python object of known type, restricted to a GIL lifetime just like PyAny.

Used: Whenever you want 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 types all 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 you should consult the pyo3::types module.

Conversions:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
use pyo3::types::PyList;
Python::with_gil(|py| -> PyResult<()> {
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()
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

Represents: 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().

Conversions:

For a Py<PyList>, the conversions are as below:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
use pyo3::types::PyList;
Python::with_gil(|py| {
let list: Py<PyList> = PyList::empty(py).into();

// To &PyList with Py::as_ref() (borrows from the Py)
let _: &PyList = list.as_ref(py);

let list_clone = list.clone(); // Because `.into_ref()` will consume `list`.
// To &PyList with Py::into_ref() (moves the pointer into PyO3's object storage)
let _: &PyList = list.into_ref(py);

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

For a #[pyclass] struct MyClass, the conversions for Py<MyClass> are below:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
Python::with_gil(|py| {
#[pyclass] struct MyClass { }
Python::with_gil(|py| -> PyResult<()> {
let my_class: Py<MyClass> = Py::new(py, MyClass { })?;

// To &PyCell<MyClass> with Py::as_ref() (borrows from the Py)
let _: &PyCell<MyClass> = my_class.as_ref(py);

let my_class_clone = my_class.clone(); // Because `.into_ref()` will consume `my_class`.
// To &PyCell<MyClass> with Py::into_ref() (moves the pointer into PyO3's object storage)
let _: &PyCell<MyClass> = my_class.into_ref(py);

let my_class = my_class_clone.clone();
// To Py<PyAny> (aka PyObject) with .into_py(py)
let _: Py<PyAny> = my_class.into_py(py);

let my_class = my_class_clone;
// To PyRef<'_, MyClass> with Py::borrow or Py::try_borrow
let _: PyRef<'_, MyClass> = my_class.try_borrow(py)?;

// To PyRefMut<'_, MyClass> with Py::borrow_mut or Py::try_borrow_mut
let _: PyRefMut<'_, MyClass> = my_class.try_borrow_mut(py)?;
Ok(())
}).unwrap();
});
}

PyCell<SomeType>

Represents: a reference to a Rust object (instance of PyClass) which is 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, PyCell<T> implements Deref<Target = PyAny>, so it also exposes all of the methods on PyAny.

Conversions:

PyCell<T> can be used to access &T and &mut T via PyRef<T> and PyRefMut<T> respectively.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass] struct MyClass { }
Python::with_gil(|py| -> PyResult<()> {
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> can also be accessed like a Python-native type.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass] struct MyClass { }
Python::with_gil(|py| -> PyResult<()> {
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()
let _: &PyAny = cell.as_ref();
Ok(())
}).unwrap();
}

PyRef<SomeType> and PyRefMut<SomeType>

Represents: reference wrapper types employed by PyCell to keep track of borrows, analog to Ref and RefMut used by RefCell.

Used: while borrowing a PyCell. They can also be used with .extract() on types like Py<T> and PyAny to get a reference quickly.

Related traits and types

PyClass

This trait marks structs defined in Rust that are also usable as Python classes, usually defined using the #[pyclass] macro.

PyNativeType

This trait marks structs that mirror native Python types, such as PyList.

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(unused)]
fn main() {
#![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(unused)]
fn main() {
#![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(unused)]
fn main() {
#![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

See cargo expand for a more elaborate 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 and Python::acquire_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-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.

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:

• Duration -> PyDelta
• FixedOffset -> PyDelta
• Utc -> PyTzInfo
• NaiveDate -> PyDate
• NaiveTime -> PyTime
• DateTime -> PyDateTime

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.

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

#![allow(unused)]
fn main() {
#[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>>,
}
}
}

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-bound, or "owned" references, and GIL-independent Py<Any> smart pointers.

GIL-bound memory

PyO3's GIL-bound, "owned references" (&PyAny etc.) 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<()> {
    let hello: &PyString = py.eval("\"Hello World!\"", None, None)?.extract()?;
    println!("Python says: {}", hello);
    Ok(())
})?;
Ok(())
}

Internally, calling Python::with_gil() or Python::acquire_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 or the GILGuard from acquire_gil() is dropped, 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 {
        let hello: &PyString = py.eval("\"Hello World!\"", None, None)?.extract()?;
        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<()> {
        let hello: &PyString = py.eval("\"Hello World!\"", None, None)?.extract()?;
        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 {
        let pool = unsafe { py.new_pool() };
        let py = pool.python();
        let hello: &PyString = py.eval("\"Hello World!\"", None, None)?.extract()?;
        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.

This behavior may change in future, see issue #1056.

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<()> {
    let hello: Py<PyString> = py.eval("\"Hello World!\"", None, None)?.extract()?;
    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| {
    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| {
    println!("Python says: {}", hello.as_ref(py));
});
// 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<()> {
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| {
    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<()> {
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| {
    println!("Python says: {}", hello.into_ref(py));
    // Memory not released yet.
    // Do more stuff...
    // Memory released here at end of `with_gil()` closure.
});
Ok(())
}

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.

Memory management

PyO3's &PyAny "owned references" and Py<PyAny> smart pointers are used to access memory stored in Python's heap. This memory sometimes lives for longer than expected because of differences in Rust and Python's memory models. See the chapter on memory management for more information.

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.

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.

See, as an example, Bazel rules to build PyO3 on Linux at https://github.com/TheButlah/rules_pyo3.

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 = "0.18.3", 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_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 = "0.18.3", 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:

#![allow(unused)]
fn main() {
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).

#![allow(unused)]
fn main() {
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(_py: Python<'_>, m: &PyModule) -> PyResult<()> {
    // A good place to install the Rust -> Python logger.
    pyo3_log::init();

    m.add_function(wrap_pyfunction!(log_something))?;
    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

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:

#![allow(unused)]
fn main() {
//! lib.rs

use pyo3::{prelude::*, wrap_pyfunction};

#[pyfunction]
fn rust_sleep(py: Python<'_>) -> PyResult<&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(py: Python<'_>, m: &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:

#![allow(unused)]
fn main() {
//! lib.rs

use pyo3::{prelude::*, wrap_pyfunction};

#[pyfunction]
fn rust_sleep(py: Python<'_>) -> PyResult<&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(py: Python<'_>, m: &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::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:

#![allow(unused)]
fn main() {
#[pyfunction]
fn await_coro(coro: &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:

#![allow(unused)]
fn main() {
#[pyfunction]
fn await_coro(callable: &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:

#![allow(unused)]
fn main() {
/// 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:

#![allow(unused)]
fn main() {
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:

#![allow(unused)]
fn main() {
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<&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::run_forever or 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"

#![allow(unused)]
fn main() {
//! lib.rs

use pyo3::{prelude::*, wrap_pyfunction};

#[pyfunction]
fn rust_sleep(py: Python<'_>) -> PyResult<&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(_py: Python<'_>, m: &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

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 = "0.18.3"

[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 = "0.18.3"

[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:

#![allow(unused)]
fn main() {
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<...>:

#![allow(unused)]
fn main() {
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 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:

#![allow(unused)]
fn main() {
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. The two easiest 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 0.17.* to 0.18

Required arguments after Option<_> arguments will no longer be automatically inferred

In #[pyfunction] and #[pymethods], if a "required" function input such as i32 came after an Option<_> input, then the Option<_> would be implicitly treated as required. (All trailing Option<_> arguments were treated as optional with a default value of None).

Starting with PyO3 0.18, this is deprecated and a future PyO3 version will require a #[pyo3(signature = (...))] option to explicitly declare the programmer's intention.

Before, x in the below example would be required to be passed from Python code:

#![allow(unused)]
fn main() {
#![allow(dead_code)]
use pyo3::prelude::*;

#[pyfunction]
fn required_argument_after_option(x: Option<i32>, y: i32) { }
}

After, specify the intended Python signature explicitly:

#![allow(unused)]
fn main() {
#![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) { }
}

__text_signature__ is now automatically generated for #[pyfunction] and #[pymethods]

The #[pyo3(text_signature = "...")] option was previously the only supported way to set the __text_signature__ attribute on generated Python functions.

PyO3 is now able to automatically populate __text_signature__ for all functions automatically based on their Rust signature (or the new #[pyo3(signature = (...))] option). These automatically-generated __text_signature__ values will currently only render ... for all default values. Many #[pyo3(text_signature = "...")] options can be removed from functions when updating to PyO3 0.18, however in cases with default values a manual implementation may still be preferred for now.

As examples:

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!(simple_function, py).unwrap();
        assert_eq!(simple.getattr("__text_signature__").unwrap().to_string(), "(a, b, c)");
        let defaulted = wrap_pyfunction!(function_with_defaults, py).unwrap();
        assert_eq!(defaulted.getattr("__text_signature__").unwrap().to_string(), "(a, b=1, c=2)");
    })
}

from 0.16.* to 0.17

Type checks have been changed for PyMapping and PySequence types

Previously the type checks for PyMapping and PySequence (implemented in PyTryFrom) used the Python C-API functions PyMapping_Check and PySequence_Check. Unfortunately these functions are not sufficient for distinguishing such types, leading to inconsistent behavior (see pyo3/pyo3#2072).

PyO3 0.17 changes these downcast checks to explicitly test if the type is a subclass of the corresponding abstract base class collections.abc.Mapping or collections.abc.Sequence. Note this requires calling into Python, which may incur a performance penalty over the previous method. If this performance penalty is a problem, you may be able to perform your own checks and use try_from_unchecked (unsafe).

Another side-effect is that a pyclass defined in Rust with PyO3 will need to be registered with the corresponding Python abstract base class for downcasting to succeed. PySequence::register and PyMapping:register have been added to make it easy to do this from Rust code. These are equivalent to calling collections.abc.Mapping.register(MappingPyClass) or collections.abc.Sequence.register(SequencePyClass) from Python.

For example, for a mapping class defined in Rust:

#![allow(unused)]
fn main() {
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
}
}

You must register the class with collections.abc.Mapping before the downcast will work:

#![allow(unused)]
fn main() {
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());
}

Note that this requirement may go away in the future when a pyclass is able to inherit from the abstract base class directly (see pyo3/pyo3#991).

The multiple-pymethods feature now requires Rust 1.62

Due to limitations in the inventory crate which the multiple-pymethods feature depends on, this feature now requires Rust 1.62. For more information see dtolnay/inventory#32.

Added impl IntoPy<Py<PyString>> for &str

This may cause inference errors.

Before:

use pyo3::prelude::*;

fn main() {
Python::with_gil(|py| {
    // Cannot infer either `Py<PyAny>` or `Py<PyString>`
    let _test = "test".into_py(py);
});
}

After, some type annotations may be necessary:

use pyo3::prelude::*;

fn main() {
Python::with_gil(|py| {
    let _test: Py<PyAny> = "test".into_py(py);
});
}

The pyproto feature is now disabled by default

In preparation for removing the deprecated #[pyproto] attribute macro in a future PyO3 version, it is now gated behind an opt-in feature flag. This also gives a slight saving to compile times for code which does not use the deprecated macro.

PyTypeObject trait has been deprecated

The PyTypeObject trait already was near-useless; almost all functionality was already on the PyTypeInfo trait, which PyTypeObject had a blanket implementation based upon. In PyO3 0.17 the final method, PyTypeObject::type_object was moved to PyTypeInfo::type_object.

To migrate, update trait bounds and imports from PyTypeObject to PyTypeInfo.

Before:

#![allow(unused)]
fn main() {
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)
}
}

After

#![allow(unused)]
fn main() {
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); });
}

impl<T, const N: usize> IntoPy<PyObject> for [T; N] now requires T: IntoPy rather than T: ToPyObject

If this leads to errors, simply implement IntoPy. Because pyclasses already implement IntoPy, you probably don't need to worry about this.

Each #[pymodule] can now only be initialized once per process

To make PyO3 modules sound in the presence of Python sub-interpreters, for now it has been necessary to explicitly disable the ability to initialize a #[pymodule] more than once in the same process. Attempting to do this will now raise an ImportError.

from 0.15.* to 0.16

Drop support for older technologies

PyO3 0.16 has increased minimum Rust version to 1.48 and minimum Python version to 3.7. This enables use of newer language features (enabling some of the other additions in 0.16) and simplifies maintenance of the project.

#[pyproto] has been deprecated

In PyO3 0.15, the #[pymethods] attribute macro gained support for implementing "magic methods" such as __str__ (aka "dunder" methods). This implementation was not quite finalized at the time, with a few edge cases to be decided upon. The existing #[pyproto] attribute macro was left untouched, because it covered these edge cases.

In PyO3 0.16, the #[pymethods] implementation has been completed and is now the preferred way to implement magic methods. To allow the PyO3 project to move forward, #[pyproto] has been deprecated (with expected removal in PyO3 0.18).

Migration from #[pyproto] to #[pymethods] is straightforward; copying the existing methods directly from the #[pyproto] trait implementation is all that is needed in most cases.

Before:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
use pyo3::class::{PyBasicProtocol, PyIterProtocol};
use pyo3::types::PyString;

#[pyclass]
struct MyClass {}

#[pyproto]
impl PyBasicProtocol 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()
    }
}
}

After

#![allow(unused)]
fn main() {
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()
    }
}
}

Removed PartialEq for object wrappers

The Python object wrappers Py and PyAny had implementations of PartialEq so that object_a == object_b would compare the Python objects for pointer equality, which corresponds to the is operator, not the == operator in Python. This has been removed in favor of a new method: use object_a.is(object_b). This also has the advantage of not requiring the same wrapper type for object_a and object_b; you can now directly compare a Py<T> with a &PyAny without having to convert.

To check for Python object equality (the Python == operator), use the new method eq().

Container magic methods now match Python behavior

In PyO3 0.15, __getitem__, __setitem__ and __delitem__ in #[pymethods] would generate only the mapping implementation for a #[pyclass]. To match the Python behavior, these methods now generate both the mapping and sequence implementations.

This means that classes implementing these #[pymethods] will now also be treated as sequences, same as a Python class would be. Small differences in behavior may result:

• PyO3 will allow instances of these classes to be cast to PySequence as well as PyMapping.
• Python will provide a default implementation of __iter__ (if the class did not have one) which repeatedly calls __getitem__ with integers (starting at 0) until an IndexError is raised.

To explain this in detail, consider the following Python class:

class ExampleContainer:

    def __len__(self):
        return 5

    def __getitem__(self, idx: int) -> int:
        if idx < 0 or idx > 5:
            raise IndexError()
        return idx

This class implements a Python sequence.

The __len__ and __getitem__ methods are also used to implement a Python mapping. In the Python C-API, these methods are not shared: the sequence __len__ and __getitem__ are defined by the sq_length and sq_item slots, and the mapping equivalents are mp_length and mp_subscript. There are similar distinctions for __setitem__ and __delitem__.

Because there is no such distinction from Python, implementing these methods will fill the mapping and sequence slots simultaneously. A Python class with __len__ implemented, for example, will have both the sq_length and mp_length slots filled.

The PyO3 behavior in 0.16 has been changed to be closer to this Python behavior by default.

wrap_pymodule! and wrap_pyfunction! now respect privacy correctly

Prior to PyO3 0.16 the wrap_pymodule! and wrap_pyfunction! macros could use modules and functions whose defining fn was not reachable according Rust privacy rules.

For example, the following code was legal before 0.16, but in 0.16 is rejected because the wrap_pymodule! macro cannot access the private_submodule function:

#![allow(unused)]
fn main() {
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(())
}
}

To fix it, make the private submodule visible, e.g. with pub or pub(crate).

#![allow(unused)]
fn main() {
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(())
}
}

from 0.14.* to 0.15

Changes in sequence indexing

For all types that take sequence indices (PyList, PyTuple and PySequence), the API has been made consistent to only take usize indices, for consistency with Rust's indexing conventions. Negative indices, which were only sporadically supported even in APIs that took isize, now aren't supported anywhere.

Further, the get_item methods now always return a PyResult instead of panicking on invalid indices. The Index trait has been implemented instead, and provides the same panic behavior as on Rust vectors.

Note that slice indices (accepted by PySequence::get_slice and other) still inherit the Python behavior of clamping the indices to the actual length, and not panicking/returning an error on out of range indices.

An additional advantage of using Rust's indexing conventions for these types is that these types can now also support Rust's indexing operators as part of a consistent API:

#![allow(unused)]
fn main() {
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]");
});
}

from 0.13.* to 0.14

auto-initialize feature is now opt-in

For projects embedding Python in Rust, PyO3 no longer automatically initializes a Python interpreter on the first call to Python::with_gil (or Python::acquire_gil) unless the auto-initialize feature is enabled.

New multiple-pymethods feature

#[pymethods] have been reworked with a simpler default implementation which removes the dependency on the inventory crate. This reduces dependencies and compile times for the majority of users.

The limitation of the new default implementation is that it cannot support multiple #[pymethods] blocks for the same #[pyclass]. If you need this functionality, you must enable the multiple-pymethods feature which will switch #[pymethods] to the inventory-based implementation.

Deprecated #[pyproto] methods

Some protocol (aka __dunder__) methods such as __bytes__ and __format__ have been possible to implement two ways in PyO3 for some time: via a #[pyproto] (e.g. PyBasicProtocol for the methods listed here), or by writing them directly in #[pymethods]. This is only true for a handful of the #[pyproto] methods (for technical reasons to do with the way PyO3 currently interacts with the Python C-API).

In the interest of having only one way to do things, the #[pyproto] forms of these methods have been deprecated.

To migrate just move the affected methods from a #[pyproto] to a #[pymethods] block.

Before:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
use pyo3::class::basic::PyBasicProtocol;

#[pyclass]
struct MyClass {}

#[pyproto]
impl PyBasicProtocol for MyClass {
    fn __bytes__(&self) -> &'static [u8] {
        b"hello, world"
    }
}
}

After:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

#[pyclass]
struct MyClass {}

#[pymethods]
impl MyClass {
    fn __bytes__(&self) -> &'static [u8] {
        b"hello, world"
    }
}
}

from 0.12.* to 0.13

Minimum Rust version increased to Rust 1.45

PyO3 0.13 makes use of new Rust language features stabilized between Rust 1.40 and Rust 1.45. If you are using a Rust compiler older than Rust 1.45, you will need to update your toolchain to be able to continue using PyO3.

Runtime changes to support the CPython limited API

In PyO3 0.13 support was added for compiling against the CPython limited API. This had a number of implications for all PyO3 users, described here.

The largest of these is that all types created from PyO3 are what CPython calls "heap" types. The specific implications of this are:

• If you wish to subclass one of these types from Rust you must mark it #[pyclass(subclass)], as you would if you wished to allow subclassing it from Python code.
• Type objects are now mutable - Python code can set attributes on them.
• __module__ on types without #[pyclass(module="mymodule")] no longer returns builtins, it now raises AttributeError.

from 0.11.* to 0.12

PyErr has been reworked

In PyO3 0.12 the PyErr type has been re-implemented to be significantly more compatible with the standard Rust error handling ecosystem. Specifically PyErr now implements Error + Send + Sync, which are the standard traits used for error types.

While this has necessitated the removal of a number of APIs, the resulting PyErr type should now be much more easier to work with. The following sections list the changes in detail and how to migrate to the new APIs.

PyErr::new and PyErr::from_type now require Send + Sync for their argument

For most uses no change will be needed. If you are trying to construct PyErr from a value that is not Send + Sync, you will need to first create the Python object and then use PyErr::from_instance.

Similarly, any types which implemented PyErrArguments will now need to be Send + Sync.

PyErr's contents are now private

It is no longer possible to access the fields .ptype, .pvalue and .ptraceback of a PyErr. You should instead now use the new methods PyErr::ptype, PyErr::pvalue and PyErr::ptraceback.

PyErrValue and PyErr::from_value have been removed

As these were part the internals of PyErr which have been reworked, these APIs no longer exist.

If you used this API, it is recommended to use PyException::new_err (see the section on Exception types).

Into<PyResult<T>> for PyErr has been removed

This implementation was redundant. Just construct the Result::Err variant directly.

Before:

#![allow(unused)]
fn main() {
let result: PyResult<()> = PyErr::new::<TypeError, _>("error message").into();
}

After (also using the new reworked exception types; see the following section):

#![allow(unused)]
fn main() {
use pyo3::{PyResult, exceptions::PyTypeError};
let result: PyResult<()> = Err(PyTypeError::new_err("error message"));
}

Exception types have been reworked

Previously exception types were zero-sized marker types purely used to construct PyErr. In PyO3 0.12, these types have been replaced with full definitions and are usable in the same way as PyAny, PyDict etc. This makes it possible to interact with Python exception objects.

The new types also have names starting with the "Py" prefix. For example, before:

#![allow(unused)]
fn main() {
let err: PyErr = TypeError::py_err("error message");
}

After:

#![allow(unused)]
fn main() {
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();
}

FromPy has been removed

To simplify the PyO3 conversion traits, the FromPy trait has been removed. Previously there were two ways to define the to-Python conversion for a type: FromPy<T> for PyObject and IntoPy<PyObject> for T.

Now there is only one way to define the conversion, IntoPy, so downstream crates may need to adjust accordingly.

Before:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
struct MyPyObjectWrapper(PyObject);

impl FromPy<MyPyObjectWrapper> for PyObject {
    fn from_py(other: MyPyObjectWrapper, _py: Python<'_>) -> Self {
        other.0
    }
}
}

After

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
struct MyPyObjectWrapper(PyObject);

impl IntoPy<PyObject> for MyPyObjectWrapper {
    fn into_py(self, _py: Python<'_>) -> PyObject {
        self.0
    }
}
}

Similarly, code which was using the FromPy trait can be trivially rewritten to use IntoPy.

Before:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
Python::with_gil(|py| {
let obj = PyObject::from_py(1.234, py);
})
}

After:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
Python::with_gil(|py| {
let obj: PyObject = 1.234.into_py(py);
})
}

PyObject is now a type alias of Py<PyAny>

This should change very little from a usage perspective. If you implemented traits for both PyObject and Py<T>, you may find you can just remove the PyObject implementation.

AsPyRef has been removed

As PyObject has been changed to be just a type alias, the only remaining implementor of AsPyRef was Py<T>. This removed the need for a trait, so the AsPyRef::as_ref method has been moved to Py::as_ref.

This should require no code changes except removing use pyo3::AsPyRef for code which did not use pyo3::prelude::*.

Before:

#![allow(unused)]
fn main() {
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);
})
}

After:

#![allow(unused)]
fn main() {
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);
})
}

from 0.10.* to 0.11

Stable Rust

PyO3 now supports the stable Rust toolchain. The minimum required version is 1.39.0.

#[pyclass] structs must now be Send or unsendable

Because #[pyclass] structs can be sent between threads by the Python interpreter, they must implement Send or declared as unsendable (by #[pyclass(unsendable)]). Note that unsendable is added in PyO3 0.11.1 and Send is always required in PyO3 0.11.0.

This may "break" some code which previously was accepted, even though it could be unsound. There can be two fixes:

1. If you think that your #[pyclass] actually must be Sendable, then let's implement Send. A common, safer way is using thread-safe types. E.g., Arc instead of Rc, Mutex instead of RefCell, and Box<dyn Send + T> instead of Box<dyn T>.

   Before:

   #![allow(unused)]
   fn main() {
   use pyo3::prelude::*;
   use std::rc::Rc;
   use std::cell::RefCell;
   
   #[pyclass]
   struct NotThreadSafe {
       shared_bools: Rc<RefCell<Vec<bool>>>,
       closure: Box<dyn Fn()>,
   }
   }
   

   After:

   #![allow(unused)]
   fn main() {
   #![allow(dead_code)]
   use pyo3::prelude::*;
   use std::sync::{Arc, Mutex};
   
   #[pyclass]
   struct ThreadSafe {
       shared_bools: Arc<Mutex<Vec<bool>>>,
       closure: Box<dyn Fn() + Send>,
   }
   }
   

   In situations where you cannot change your #[pyclass] to automatically implement Send (e.g., when it contains a raw pointer), you can use unsafe impl Send. In such cases, care should be taken to ensure the struct is actually thread safe. See the Rustonomicon for more.

2. If you think that your #[pyclass] should not be accessed by another thread, you can use unsendable flag. A class marked with unsendable panics when accessed by another thread, making it thread-safe to expose an unsendable object to the Python interpreter.

   Before:

   #![allow(unused)]
   fn main() {
   use pyo3::prelude::*;
   
   #[pyclass]
   struct Unsendable {
       pointers: Vec<*mut std::os::raw::c_char>,
   }
   }
   

   After:

   #![allow(unused)]
   fn main() {
   #![allow(dead_code)]
   use pyo3::prelude::*;
   
   #[pyclass(unsendable)]
   struct Unsendable {
       pointers: Vec<*mut std::os::raw::c_char>,
   }
   }
   

All PyObject and Py<T> methods now take Python as an argument

Previously, a few methods such as Object::get_refcnt did not take Python as an argument (to ensure that the Python GIL was held by the current thread). Technically, this was not sound. To migrate, just pass a py argument to any calls to these methods.

Before:

#![allow(unused)]
fn main() {
pyo3::Python::with_gil(|py| {
py.None().get_refcnt();
})
}

After:

#![allow(unused)]
fn main() {
pyo3::Python::with_gil(|py| {
py.None().get_refcnt(py);
})
}

from 0.9.* to 0.10

ObjectProtocol is removed

All methods are moved to PyAny. And since now all native types (e.g., PyList) implements Deref<Target=PyAny>, all you need to do is remove ObjectProtocol from your code. Or if you use ObjectProtocol by use pyo3::prelude::*, you have to do nothing.

Before:

#![allow(unused)]
fn main() {
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);
})
}

After:

#![allow(unused)]
fn main() {
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);
})
}

No #![feature(specialization)] in user code

While PyO3 itself still requires specialization and nightly Rust, now you don't have to use #![feature(specialization)] in your crate.

from 0.8.* to 0.9

#[new] interface

PyRawObject is now removed and our syntax for constructors has changed.

Before:

#![allow(unused)]
fn main() {
#[pyclass]
struct MyClass {}

#[pymethods]
impl MyClass {
    #[new]
    fn new(obj: &PyRawObject) {
        obj.init(MyClass {})
    }
}
}

After:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
struct MyClass {}

#[pymethods]
impl MyClass {
    #[new]
    fn new() -> Self {
        MyClass {}
    }
}
}

Basically you can return Self or Result<Self> directly. For more, see the constructor section of this guide.

PyCell

PyO3 0.9 introduces PyCell, which is a RefCell-like object wrapper for ensuring Rust's rules regarding aliasing of references are upheld. For more detail, see the Rust Book's section on Rust's rules of references

For #[pymethods] or #[pyfunction]s, your existing code should continue to work without any change. Python exceptions will automatically be raised when your functions are used in a way which breaks Rust's rules of references.

Here is an example.

#![allow(unused)]
fn main() {
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 = PyCell::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'
    ");
})
}

Names has a merge method, which takes &mut self and another argument of type &mut Self. Given this #[pyclass], calling names.merge(names) in Python raises a PyBorrowMutError exception, since it requires two mutable borrows of names.

However, for #[pyproto] and some functions, you need to manually fix the code.

Object creation

In 0.8 object creation was done with PyRef::new and PyRefMut::new. In 0.9 these have both been removed. To upgrade code, please use PyCell::new instead. If you need PyRef or PyRefMut, just call .borrow() or .borrow_mut() on the newly-created PyCell.

Before:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
struct MyClass {}
Python::with_gil(|py| {
let obj_ref = PyRef::new(py, MyClass {}).unwrap();
})
}

After:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;
#[pyclass]
struct MyClass {}
Python::with_gil(|py| {
let obj = PyCell::new(py, MyClass {}).unwrap();
let obj_ref = obj.borrow();
})
}

Object extraction

For PyClass types T, &T and &mut T no longer have FromPyObject implementations. Instead you should extract PyRef<T> or PyRefMut<T>, respectively. If T implements Clone, you can extract T itself. In addition, you can also extract &PyCell<T>, though you rarely need it.

Before:

let obj: &PyAny = create_obj();
let obj_ref: &MyClass = obj.extract().unwrap();
let obj_ref_mut: &mut MyClass = obj.extract().unwrap();

After:

#![allow(unused)]
fn main() {
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();
})
}

#[pyproto]

Most of the arguments to methods in #[pyproto] impls require a FromPyObject implementation. So if your protocol methods take &T or &mut T (where T: PyClass), please use PyRef or PyRefMut instead.

Before:

#![allow(unused)]
fn main() {
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 })
    }
}
}

After:

#![allow(unused)]
fn main() {
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 })
    }
}
}

PyO3 and rust-cpython

PyO3 began as fork of rust-cpython when rust-cpython wasn't maintained. Over time PyO3 has become fundamentally different from rust-cpython.

Macros

While rust-cpython has a macro_rules! based DSL for declaring modules and classes, PyO3 uses proc macros. PyO3 also doesn't change your struct and functions so you can still use them as normal Rust functions.

rust-cpython

py_class!(class MyClass |py| {
    data number: i32;
    def __new__(_cls, arg: i32) -> PyResult<MyClass> {
        MyClass::create_instance(py, arg)
    }
    def half(&self) -> PyResult<i32> {
        Ok(self.number(py) / 2)
    }
});

PyO3

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

#[pyclass]
struct MyClass {
    num: u32,
}

#[pymethods]
impl MyClass {
    #[new]
    fn new(num: u32) -> Self {
        MyClass { num }
    }

    fn half(&self) -> PyResult<u32> {
        Ok(self.num / 2)
    }
}
}

Ownership and lifetimes

While in rust-cpython you always own Python objects, PyO3 allows efficient borrowed objects and most APIs work with references.

Here is an example of the PyList API:

rust-cpython

impl PyList {

   fn new(py: Python<'_>) -> PyList {...}

   fn get_item(&self, py: Python<'_>, index: isize) -> PyObject {...}
}

PyO3

impl PyList {

   fn new(py: Python<'_>) -> &PyList {...}

   fn get_item(&self, index: isize) -> &PyAny {...}
}

In PyO3, all object references are bounded by the GIL lifetime. So owned Python objects are not required, and it is safe to have functions like fn py<'p>(&'p self) -> Python<'p> {}.

Error handling

rust-cpython requires a Python parameter for constructing a PyErr, so error handling ergonomics is pretty bad. It is not possible to use ? with Rust errors.

PyO3 on other hand does not require Python for constructing a PyErr, it is only required if you want to raise an exception in Python with the PyErr::restore() method. Due to various std::convert::From<E> for PyErr implementations for Rust standard error types E, propagating ? is supported automatically.

Using in Python a Rust function with trait bounds

PyO3 allows for easy conversion from Rust to Python for certain functions and classes (see the conversion table. However, it is not always straightforward to convert Rust code that requires a given trait implementation as an argument.

This tutorial explains how to convert a Rust function that takes a trait as argument for use in Python with classes implementing the same methods as the trait.

Why is this useful?

Pros

• Make your Rust code available to Python users
• Code complex algorithms in Rust with the help of the borrow checker

Cons

• Not as fast as native Rust (type conversion has to be performed and one part of the code runs in Python)
• You need to adapt your code to expose it

Example

Let's work with the following basic example of an implementation of a optimization solver operating on a given model.

Let's say we have a function solve that operates on a model and mutates its state. The argument of the function can be any model that implements the Model trait :

#![allow(unused)]
fn main() {
#![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");
}
}

Let's assume we have the following constraints:

• We cannot change that code as it runs on many Rust models.
• We also have many Python models that cannot be solved as this solver is not available in that language. Rewriting it in Python would be cumbersome and error-prone, as everything is already available in Rust.

How could we expose this solver to Python thanks to PyO3 ?

Implementation of the trait bounds for the Python class

If a Python class implements the same three methods as the Model trait, it seems logical it could be adapted to use the solver. However, it is not possible to pass a PyObject to it as it does not implement the Rust trait (even if the Python model has the required methods).

In order to implement the trait, we must write a wrapper around the calls in Rust to the Python model. The method signatures must be the same as the trait, keeping in mind that the Rust trait cannot be changed for the purpose of making the code available in Python.

The Python model we want to expose is the following one, which already contains all the required methods:

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

The following wrapper will call the Python model from Rust, using a struct to hold the model as a PyAny object:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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| {
            let values: Vec<f64> = var.clone();
            let list: PyObject = values.into_py(py);
            let py_model = self.model.as_ref(py);
            py_model
                .call_method("set_variables", (list,), None)
                .unwrap();
        })
    }

    fn get_results(&self) -> Vec<f64> {
        println!("Rust calling Python to get the results");
        Python::with_gil(|py| {
            self.model
                .as_ref(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
                .as_ref(py)
                .call_method("compute", (), None)
                .unwrap();
        })
    }
}
}

Now that this bit is implemented, let's expose the model wrapper to Python. Let's add the PyO3 annotations and add a constructor:

#![allow(unused)]
fn main() {
#![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(_py: Python<'_>, m: &PyModule) -> PyResult<()> {
    m.add_class::<UserModel>()?;
    Ok(())
}

#[pymethods]
impl UserModel {
    #[new]
    pub fn new(model: Py<PyAny>) -> Self {
        UserModel { model }
    }
}
}

Now we add the PyO3 annotations to the trait implementation:

#[pymethods]
impl Model for UserModel {
    // the previous trait implementation
}

However, the previous code will not compile. The compilation error is the following one: error: #[pymethods] cannot be used on trait impl blocks

That's a bummer! However, we can write a second wrapper around these functions to call them directly. This wrapper will also perform the type conversions between Python and Rust.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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| {
         let values: Vec<f64> = var.clone();
         let list: PyObject = values.into_py(py);
         let py_model = self.model.as_ref(py);
         py_model
             .call_method("set_variables", (list,), None)
             .unwrap();
     })
 }

 fn get_results(&self) -> Vec<f64> {
     println!("Rust calling Python to get the results");
     Python::with_gil(|py| {
         self.model
             .as_ref(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
             .as_ref(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)
    }
}
}

This wrapper handles the type conversion between the PyO3 requirements and the trait. In order to meet PyO3 requirements, this wrapper must:

• return an object of type PyResult
• use only values, not references in the method signatures

Let's run the file python file:

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())

This outputs:

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]

We have now successfully exposed a Rust model that implements the Model trait to Python!

We will now expose the solve function, but before, let's talk about types errors.

Type errors in Python

What happens if you have type errors when using Python and how can you improve the error messages?

Wrong types in Python function arguments

Let's assume in the first case that you will use in your Python file my_rust_model.set_variables(2.0) instead of my_rust_model.set_variables([2.0]).

The Rust signature expects a vector, which corresponds to a list in Python. What happens if instead of a vector, we pass a single value ?

At the execution of Python, we get :

File "main.py", line 15, in <module>
   my_rust_model.set_variables(2)
TypeError

It is a type error and Python points to it, so it's easy to identify and solve.

Wrong types in Python method signatures

Let's assume now that the return type of one of the methods of our Model class is wrong, for example the get_results method that is expected to return a Vec<f64> in Rust, a list in Python.

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

This call results in the following panic:

pyo3_runtime.PanicException: called `Result::unwrap()` on an `Err` value: PyErr { type: Py(0x10dcf79f0, PhantomData) }

This error code is not helpful for a Python user that does not know anything about Rust, or someone that does not know PyO3 was used to interface the Rust code.

However, as we are responsible for making the Rust code available to Python, we can do something about it.

The issue is that we called unwrap anywhere we could, and therefore any panic from PyO3 will be directly forwarded to the end user.

Let's modify the code performing the type conversion to give a helpful error message to the Python user:

We used in our get_results method the following call that performs the type conversion:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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
                .as_ref(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| {
            let values: Vec<f64> = var.clone();
            let list: PyObject = values.into_py(py);
            let py_model = self.model.as_ref(py);
            py_model
                .call_method("set_variables", (list,), None)
                .unwrap();
        })
    }

    fn compute(&mut self) {
        println!("Rust calling Python to perform the computation");
        Python::with_gil(|py| {
            self.model
                .as_ref(py)
                .call_method("compute", (), None)
                .unwrap();
        })
    }
}
}

Let's break it down in order to perform better error handling:

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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: &PyAny = self
                .model
                .as_ref(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 values: Vec<f64> = var.clone();
            let list: PyObject = values.into_py(py);
            let py_model = self.model.as_ref(py);
            py_model
                .call_method("set_variables", (list,), None)
                .unwrap();
        })
    }

    fn compute(&mut self) {
        println!("Rust calling Python to perform the computation");
        Python::with_gil(|py| {
            self.model
                .as_ref(py)
                .call_method("compute", (), None)
                .unwrap();
        })
    }
}
}

By doing so, you catch the result of the Python computation and check its type in order to be able to deliver a better error message before performing the unwrapping.

Of course, it does not cover all the possible wrong outputs: the user could return a list of strings instead of a list of floats. In this case, a runtime panic would still occur due to PyO3, but with an error message much more difficult to decipher for non-rust user.

It is up to the developer exposing the rust code to decide how much effort to invest into Python type error handling and improved error messages.

The final code

Now let's expose the solve() function to make it available from Python.

It is not possible to directly expose the solve function to Python, as the type conversion cannot be performed. It requires an object implementing the Model trait as input.

However, the UserModel already implements this trait. Because of this, we can write a function wrapper that takes the UserModel--which has already been exposed to Python--as an argument in order to call the core function solve.

It is also required to make the struct public.

#![allow(unused)]
fn main() {
use pyo3::prelude::*;

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(_py: Python<'_>, m: &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| {
            let values: Vec<f64> = var.clone();
            let list: PyObject = values.into_py(py);
            let py_model = self.model.as_ref(py);
            py_model
                .call_method("set_variables", (list,), None)
                .unwrap();
        })
    }

    fn get_results(&self) -> Vec<f64> {
        println!("Get results from Rust calling Python");
        Python::with_gil(|py| {
            let py_result: &PyAny = self
                .model
                .as_ref(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
                .as_ref(py)
                .call_method("compute", (), None)
                .unwrap();
        })
    }
}
}

Typing and IDE hints for you Python package

PyO3 provides an easy to use interface to code native Python libraries in Rust. The accompanying Maturin allows you to build and publish them as a package. Yet, for the better user experience, Python libraries should provide typing hints and documentation for all public entities, so that IDEs can show them during development and type analyzing tools such as mypy can use them to properly verify the code.

Currently the best solution for the problem is to manually maintain *.pyi files and ship them along with the package.

There is a sketch of a roadmap towards completing the experimental-inspect feature which may eventually lead to automatic type annotations generated by PyO3. This needs more testing and implementation, please see issue #2454.

Introduction to pyi files

pyi files (an abbreviation for Python Interface) are called "stub files" in most of the documentation related to them. A very good definition of what it is can be found in old MyPy documentation:

A stubs file only contains a description of the public interface of the module without any implementations.

There is also extensive documentation on type stubs on the offical Python typing documentation.

Most Python developers probably already encountered them when trying to use their IDE's "Go to Definition" function on any builtin type. For example, the definitions of a few standard exceptions look like this:

class BaseException(object):
    args: Tuple[Any, ...]
    __cause__: BaseException | None
    __context__: BaseException | None
    __suppress_context__: bool
    __traceback__: TracebackType | None
    def __init__(self, *args: object) -> None: ...
    def __str__(self) -> str: ...
    def __repr__(self) -> str: ...
    def with_traceback(self: _TBE, tb: TracebackType | None) -> _TBE: ...

class SystemExit(BaseException):
    code: int

class Exception(BaseException): ...

class StopIteration(Exception):
    value: Any

As we can see, those are not full definitions containing implementation, but just a description of interface. It is usually all that the user of the library needs.

What do the PEPs say?

At the time of writing this documentation, the pyi files are referenced in three PEPs.

PEP8 - Style Guide for Python Code - #Function Annotations (last point) recommends all third party library creators to provide stub files as the source of knowledge about the package for type checker tools.

(...) it is expected that users of third party library packages may want to run type checkers over those packages. For this purpose PEP 484 recommends the use of stub files: .pyi files that are read by the type checker in preference of the corresponding .py files. (...)

PEP484 - Type Hints - #Stub Files defines stub files as follows.

Stub files are files containing type hints that are only for use by the type checker, not at runtime.

It contains a specification for them (highly recommended reading, since it contains at least one thing that is not used in normal Python code) and also some general information about where to store the stub files.

PEP561 - Distributing and Packaging Type Information describes in detail how to build packages that will enable type checking. In particular it contains information about how the stub files must be distributed in order for type checkers to use them.

How to do it?

PEP561 recognizes three ways of distributing type information:

• inline - the typing is placed directly in source (py) files;
• separate package with stub files - the typing is placed in pyi files distributed in their own, separate package;
• in-package stub files - the typing is placed in pyi files distributed in the same package as source files.

The first way is tricky with PyO3 since we do not have py files. When it has been investigated and necessary changes are implemented, this document will be updated.

The second way is easy to do, and the whole work can be fully separated from the main library code. The example repo for the package with stub files can be found in PEP561 references section: Stub package repository

The third way is described below.

Including pyi files in your PyO3/Maturin build package

When source files are in the same package as stub files, they should be placed next to each other. We need a way to do that with Maturin. Also, in order to mark our package as typing-enabled we need to add an empty file named py.typed to the package.

If you do not have other Python files

If you do not need to add any other Python files apart from pyi to the package, Maturin provides a way to do most of the work for you. As documented in the Maturin Guide, the only thing you need to do is to create a stub file for your module named <module_name>.pyi in your project root and Maturin will do the rest.

my-rust-project/
├── Cargo.toml
├── my_project.pyi  # <<< add type stubs for Rust functions in the my_project module here
├── pyproject.toml
└── src
    └── lib.rs

For an example pyi file see the my_project.pyi content section.

If you need other Python files

If you need to add other Python files apart from pyi to the package, you can do it also, but that requires some more work. Maturin provides an easy way to add files to a package (documentation). You just need to create a folder with the name of your module next to the Cargo.toml file (for customization see documentation linked above).

The folder structure would be:

my-project
├── Cargo.toml
├── my_project
│   ├── __init__.py
│   ├── my_project.pyi
│   ├── other_python_file.py
│   └── py.typed
├── pyproject.toml
├── Readme.md
└── src
    └── lib.rs

Let's go a little bit more into details regarding the files inside the package folder.

__init__.py content

As we now specify our own package content, we have to provide the __init__.py file, so the folder is treated as a package and we can import things from it. We can always use the same content that Maturin creates for us if we do not specify a Python source folder. For PyO3 bindings it would be:

from .my_project import *

That way everything that is exposed by our native module can be imported directly from the package.

py.typed requirement

As stated in PEP561:

Package maintainers who wish to support type checking of their code MUST add a marker file named py.typed to their package supporting typing. This marker applies recursively: if a top-level package includes it, all its sub-packages MUST support type checking as well.

If we do not include that file, some IDEs might still use our pyi files to show hints, but the type checkers might not. MyPy will raise an error in this situation:

error: Skipping analyzing "my_project": found module but no type hints or library stubs

The file is just a marker file, so it should be empty.

my_project.pyi content

Our module stub file. This document does not aim at describing how to write them, since you can find a lot of documentation on it, starting from the already quoted PEP484.

The example can look like this:

class Car:
    """
    A class representing a car.

    :param body_type: the name of body type, e.g. hatchback, sedan
    :param horsepower: power of the engine in horsepower
    """
    def __init__(self, body_type: str, horsepower: int) -> None: ...

    @classmethod
    def from_unique_name(cls, name: str) -> 'Car':
        """
        Creates a Car based on unique name

        :param name: model name of a car to be created
        :return: a Car instance with default data
        """

    def best_color(self) -> str:
        """
        Gets the best color for the car.

        :return: the name of the color our great algorithm thinks is the best for this car
        """

Changelog

All notable changes to this project will be documented in this file. For help with updating to new PyO3 versions, please see the migration guide.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

To see unreleased changes, please see the CHANGELOG on the main branch guide.

0.18.3 - 2023-04-13

Added

• Add GILProtected<T> to mediate concurrent access to a value using Python's global interpreter lock (GIL). #2975
• Support PyASCIIObject / PyUnicode and associated methods on big-endian architectures. #3015
• Add FFI definition _PyDict_Contains_KnownHash() for CPython 3.10 and up. #3088

Fixed

• Fix compile error for #[pymethods] and #[pyfunction] called "output". #3022
• Fix compile error in generated code for magic methods implemented as a #[staticmethod]. #3055
• Fix is_instance for PyDateTime (would incorrectly check for a PyDate). #3071
• Fix upstream deprecation of PyUnicode_InternImmortal since Python 3.10. #3071

0.18.2 - 2023-03-24

Packaging

• Disable default features of chrono to avoid depending on time v0.1.x. #2939

Added

• Implement IntoPy<PyObject>, ToPyObject and FromPyObject for Cow<[u8]> to efficiently handle both bytes and bytearray objects. #2899
• Implement IntoPy<PyObject>, ToPyObject and FromPyObject for Cell<T>. #3014
• Add PyList::to_tuple(), as a convenient and efficient conversion from lists to tuples. #3042
• Add PyTuple::to_list(), as a convenient and efficient conversion from tuples to lists. #3044

Changed

• Optimize PySequence conversion for list and tuple inputs. #2944
• Improve exception raised when creating #[pyclass] type object fails during module import. #2947
• Optimize PyMapping conversion for dict inputs. #2954
• Allow create_exception! to take a dotted.module to place the exception in a submodule. #2979

Fixed

• Fix a reference counting race condition affecting PyObjects cloned in allow_threads blocks. #2952
• Fix clippy::redundant_closure lint on default arguments in #[pyo3(signature = (...))] annotations. #2990
• Fix non_snake_case lint on generated code in #[pyfunction] macro. #2993
• Fix some FFI definitions for the upcoming PyPy 3.10 release. #3031

0.18.1 - 2023-02-07

Added

• Add PyErr::write_unraisable(). #2889
• Add Python::Ellipsis() and PyAny::is_ellipsis() methods. #2911
• Add PyDict::update() and PyDict::update_if_missing() methods. #2912

Changed

• FFI definition PyIter_Check on CPython 3.7 is now implemented as hasattr(type(obj), "__next__"), which works correctly on all platforms and adds support for abi3. #2914
• Warn about unknown config keys in PYO3_CONFIG_FILE instead of denying. #2926

Fixed

• Send errors returned by __releasebuffer__ to sys.unraisablehook rather than causing SystemError. #2886
• Fix downcast to PyIterator succeeding for Python classes which did not implement __next__. #2914
• Fix segfault in __traverse__ when visiting None fields of Option<T: AsPyPointer>. #2921
• Fix #[pymethods(crate = "...")] option being ignored. #2923
• Link against pythonXY_d.dll for debug Python builds on Windows. #2937

0.18.0 - 2023-01-17

Packaging

• Relax indexmap optional depecency to allow >= 1.6, < 2. [#2849](https://github.com/PyO3/pyo3/pull/2849
• Relax hashbrown optional dependency to allow >= 0.9, < 0.14. #2875
• Update memoffset dependency to 0.8. #2875

Added

• Add GILOnceCell::get_or_try_init for fallible GILOnceCell initialization. #2398
• Add experimental feature experimental-inspect with type_input() and type_output() helpers to get the Python type of any Python-compatible object. #2490 #2882
• The #[pyclass] macro can now take get_all and set_all to create getters and setters for every field. #2692
• Add #[pyo3(signature = (...))] option for #[pyfunction] and #[pymethods]. #2702
• pyo3-build-config: rebuild when PYO3_ENVIRONMENT_SIGNATURE environment variable value changes. #2727
• Add conversions between non-zero int types in std::num and Python int. #2730
• Add Py::downcast() as a companion to PyAny::downcast(), as well as downcast_unchecked() for both types. #2734
• Add types for all built-in Warning classes as well as PyErr::warn_explicit. #2742
• Add abi3-py311 feature. #2776
• Add FFI definition _PyErr_ChainExceptions() for CPython. #2788
• Add FFI definitions PyVectorcall_NARGS and PY_VECTORCALL_ARGUMENTS_OFFSET for PyPy 3.8 and up. #2811
• Add PyList::get_item_unchecked for PyPy. #2827

Changed

• PyO3's macros now emit a much nicer error message if function return values don't implement the required trait(s). #2664
• Use a TypeError, rather than a ValueError, when refusing to treat a str as a Vec. #2685
• Change PyCFunction::new_closure to take name and doc arguments. #2686
• PyType::is_subclass, PyErr::is_instance and PyAny::is_instance now take &PyAny instead of &PyType arguments, so that they work with objects that pretend to be types using __subclasscheck__ and __instancecheck__. #2695
• Deprecate #[args] attribute and passing "args" specification directly to #[pyfunction] in favor of the new #[pyo3(signature = (...))] option. #2702
• Deprecate required arguments after Option<T> arguments to #[pyfunction] and #[pymethods] without also using #[pyo3(signature)] to specify whether the arguments should be required or have defaults. #2703
• Change #[pyfunction] and #[pymethods] to use a common call "trampoline" to slightly reduce generated code size and compile times. #2705
• PyAny::cast_as() and Py::cast_as() are now deprecated in favor of PyAny::downcast() and the new Py::downcast(). #2734
• Relax lifetime bounds on PyAny::downcast(). #2734
• Automatically generate __text_signature__ for all Python functions created using #[pyfunction] and #[pymethods]. #2784
• Accept any iterator in PySet::new and PyFrozenSet::new. #2795
• Mixing #[cfg(...)] and #[pyo3(...)] attributes on #[pyclass] struct fields will now work. #2796
• Re-enable PyFunction on when building for abi3 or PyPy. #2838
• Improve derive(FromPyObject) to use intern! when applicable for #[pyo3(item)]. #2879

Removed

• Remove the deprecated pyproto feature, #[pyproto] macro, and all accompanying APIs. #2587
• Remove all functionality deprecated in PyO3 0.16. #2843

Fixed

• Disable PyModule::filename on PyPy. #2715
• PyCodeObject is now once again defined with fields on Python 3.7. #2726
• Raise a TypeError if #[new] pymethods with no arguments receive arguments when called from Python. #2749
• Use the NOARGS argument calling convention for methods that have a single py: Python argument (as a performance optimization). #2760
• Fix truncation of isize values to c_long in PySlice::new. #2769
• Fix soundness issue with FFI definition PyUnicodeDecodeError_Create on PyPy leading to indeterminate behavior (typically a TypeError). #2772
• Allow functions taking **kwargs to accept keyword arguments which share a name with a positional-only argument (as permitted by PEP 570). #2800
• Fix unresolved symbol for PyObject_Vectorcall on PyPy 3.9 and up. #2811
• Fix memory leak in PyCFunction::new_closure. #2842

0.17.3 - 2022-11-01

Packaging

• Support Python 3.11. (Previous versions of PyO3 0.17 have been tested against Python 3.11 release candidates and are expected to be compatible, this is the first version tested against Python 3.11.0.) #2708

Added

• Implemented ExactSizeIterator for PyListIterator, PyDictIterator, PySetIterator and PyFrozenSetIterator. #2676

Fixed

• Fix regression of impl FromPyObject for [T; N] no longer accepting types passing PySequence_Check, e.g. NumPy arrays, since version 0.17.0. This the same fix that was applied impl FromPyObject for Vec<T> in version 0.17.1 extended to fixed-size arrays. #2675
• Fix UB in FunctionDescription::extract_arguments_fastcall due to creating slices from a null pointer. #2687

0.17.2 - 2022-10-04

Packaging

• Added optional chrono feature to convert chrono types into types in the datetime module. #2612

Added

• Add support for num-bigint feature on PyPy. #2626

Fixed

• Correctly implement __richcmp__ for enums, fixing __ne__ returning always returning True. #2622
• Fix compile error since 0.17.0 with Option<&SomePyClass> argument with a default. #2630
• Fix regression of impl FromPyObject for Vec<T> no longer accepting types passing PySequence_Check, e.g. NumPy arrays, since 0.17.0. #2631

0.17.1 - 2022-08-28

Fixed

• Fix visibility of PyDictItems, PyDictKeys, and PyDictValues types added in PyO3 0.17.0.
• Fix compile failure when using #[pyo3(from_py_with = "...")] attribute on an argument of type Option<T>. #2592
• Fix clippy redundant-closure lint on **kwargs arguments for #[pyfunction] and #[pymethods]. #2595

0.17.0 - 2022-08-23

Packaging

• Update inventory dependency to 0.3 (the multiple-pymethods feature now requires Rust 1.62 for correctness). #2492

Added

• Add timezone_utc. #1588
• Implement ToPyObject for [T; N]. #2313
• Add PyDictKeys, PyDictValues and PyDictItems Rust types. #2358
• Add append_to_inittab. #2377
• Add FFI definition PyFrame_GetCode. #2406
• Add PyCode and PyFrame high level objects. #2408
• Add FFI definitions Py_fstring_input, sendfunc, and _PyErr_StackItem. #2423
• Add PyDateTime::new_with_fold, PyTime::new_with_fold, PyTime::get_fold, and PyDateTime::get_fold for PyPy. #2428
• Add #[pyclass(frozen)]. #2448
• Accept #[pyo3(name)] on enum variants. #2457
• Add CompareOp::matches to implement __richcmp__ as the result of a Rust std::cmp::Ordering comparison. #2460
• Add PySuper type. #2486
• Support PyPy on Windows with the generate-import-lib feature. #2506
• Add FFI definitions Py_EnterRecursiveCall and Py_LeaveRecursiveCall. #2511
• Add PyDict::get_item_with_error. #2536
• Add #[pyclass(sequence)] option. #2567

Changed

• Change datetime constructors taking a tzinfo to take Option<&PyTzInfo> instead of Option<&PyObject>: PyDateTime::new, PyDateTime::new_with_fold, PyTime::new, and PyTime::new_with_fold. #1588
• Move PyTypeObject::type_object method to the PyTypeInfo trait, and deprecate the PyTypeObject trait. #2287
• Methods of Py and PyAny now accept impl IntoPy<Py<PyString>> rather than just &str to allow use of the intern! macro. #2312
• Change the deprecated pyproto feature to be opt-in instead of opt-out. #2322
• Emit better error messages when #[pyfunction] return types do not implement IntoPy. #2326
• Require T: IntoPy for impl<T, const N: usize> IntoPy<PyObject> for [T; N] instead of T: ToPyObject. #2326
• Deprecate the ToBorrowedObject trait. #2333
• Iterators over PySet and PyDict will now panic if the underlying collection is mutated during the iteration. #2380
• Iterators over PySet and PyDict will now panic if the underlying collection is mutated during the iteration. #2380
• Allow #[classattr] methods to be fallible. #2385
• Prevent multiple #[pymethods] with the same name for a single #[pyclass]. #2399
• Fixup lib_name when using PYO3_CONFIG_FILE. #2404
• Add a message to the ValueError raised by the #[derive(FromPyObject)] implementation for a tuple struct. #2414
• Allow #[classattr] methods to take Python argument. #2456
• Rework PyCapsule type to resolve soundness issues: #2485
  • PyCapsule::new and PyCapsule::new_with_destructor now take name: Option<CString> instead of &CStr.
  • The destructor F in PyCapsule::new_with_destructor must now be Send.
  • PyCapsule::get_context deprecated in favor of PyCapsule::context which doesn't take a py: Python<'_> argument.
  • PyCapsule::set_context no longer takes a py: Python<'_> argument.
  • PyCapsule::name now returns PyResult<Option<&CStr>> instead of &CStr.
• FromPyObject::extract for Vec<T> no longer accepts Python str inputs. #2500
• Ensure each #[pymodule] is only initialized once. #2523
• pyo3_build_config::add_extension_module_link_args now also emits linker arguments for wasm32-unknown-emscripten. #2538
• Type checks for PySequence and PyMapping now require inputs to inherit from (or register with) collections.abc.Sequence and collections.abc.Mapping respectively. #2477
• Disable PyFunction on when building for abi3 or PyPy. #2542
• Deprecate Python::acquire_gil. #2549

Removed

• Remove all functionality deprecated in PyO3 0.15. #2283
• Make the Dict, WeakRef and BaseNativeType members of the PyClass private implementation details. #2572

Fixed

• Enable incorrectly disabled FFI definition PyThreadState_DeleteCurrent. #2357
• Fix wrap_pymodule interactions with name resolution rules: it no longer "sees through" glob imports of use submodule::* when submodule::submodule is a #[pymodule]. #2363
• Correct FFI definition PyEval_EvalCodeEx to take *const *mut PyObject array arguments instead of *mut *mut PyObject. #2368
• Fix "raw-ident" structs (e.g. #[pyclass] struct r#RawName) incorrectly having r# at the start of the class name created in Python. #2395
• Correct FFI definition Py_tracefunc to be unsafe extern "C" fn (was previously safe). #2407
• Fix compile failure with #[pyo3(from_py_with = "...")] annotations on a field in a #[derive(FromPyObject)] struct. #2414
• Fix FFI definitions _PyDateTime_BaseTime and _PyDateTime_BaseDateTime lacking leading underscores in their names. #2421
• Remove FFI definition PyArena on Python 3.10 and up. #2421
• Fix FFI definition PyCompilerFlags missing member cf_feature_version on Python 3.8 and up. #2423
• Fix FFI definition PyAsyncMethods missing member am_send on Python 3.10 and up. #2423
• Fix FFI definition PyGenObject having multiple incorrect members on various Python versions. #2423
• Fix FFI definition PySyntaxErrorObject missing members end_lineno and end_offset on Python 3.10 and up. #2423
• Fix FFI definition PyHeapTypeObject missing member ht_module on Python 3.9 and up. #2423
• Fix FFI definition PyFrameObject having multiple incorrect members on various Python versions. #2424 #2434
• Fix FFI definition PyTypeObject missing deprecated field tp_print on Python 3.8. #2428
• Fix FFI definitions PyDateTime_CAPI. PyDateTime_Date, PyASCIIObject, PyBaseExceptionObject, PyListObject, and PyTypeObject on PyPy. #2428
• Fix FFI definition _inittab field initfunc typo'd as initfun. #2431
• Fix FFI definitions _PyDateTime_BaseTime and _PyDateTime_BaseDateTime incorrectly having fold member. #2432
• Fix FFI definitions PyTypeObject. PyHeapTypeObject, and PyCFunctionObject having incorrect members on PyPy 3.9. #2433
• Fix FFI definition PyGetSetDef to have *const c_char for doc member (not *mut c_char). #2439
• Fix #[pyo3(from_py_with = "...")] being ignored for 1-element tuple structs and transparent structs. #2440
• Use memoffset to avoid UB when computing PyCell layout. #2450
• Fix incorrect enum names being returned by the generated repr for enums renamed by #[pyclass(name = "...")] #2457
• Fix PyObject_CallNoArgs incorrectly being available when building for abi3 on Python 3.9. #2476
• Fix several clippy warnings generated by #[pyfunction] arguments. #2503

0.16.6 - 2022-08-23

Changed

• Fix soundness issues with PyCapsule type with select workarounds. Users are encourage to upgrade to PyO3 0.17 at their earliest convenience which contains API breakages which fix the issues in a long-term fashion. #2522
  • PyCapsule::new and PyCapsule::new_with_destructor now take ownership of a copy of the name to resolve a possible use-after-free.
  • PyCapsule::name now returns an empty CStr instead of dereferencing a null pointer if the capsule has no name.
  • The destructor F in PyCapsule::new_with_destructor will never be called if the capsule is deleted from a thread other than the one which the capsule was created in (a warning will be emitted).
• Panics during drop of panic payload caught by PyO3 will now abort. #2544

0.16.5 - 2022-05-15

Added

• Add an experimental generate-import-lib feature to support auto-generating non-abi3 python import libraries for Windows targets. #2364
• Add FFI definition Py_ExitStatusException. #2374

Changed

• Deprecate experimental generate-abi3-import-lib feature in favor of the new generate-import-lib feature. #2364

Fixed

• Added missing warn_default_encoding field to PyConfig on 3.10+. The previously missing field could result in incorrect behavior or crashes. #2370
• Fixed order of pathconfig_warnings and program_name fields of PyConfig on 3.10+. Previously, the order of the fields was swapped and this could lead to incorrect behavior or crashes. #2370

0.16.4 - 2022-04-14

Added

• Add PyTzInfoAccess trait for safe access to time zone information. #2263
• Add an experimental generate-abi3-import-lib feature to auto-generate python3.dll import libraries for Windows. #2282
• Add FFI definitions for PyDateTime_BaseTime and PyDateTime_BaseDateTime. #2294

Changed

• Improved performance of failing calls to FromPyObject::extract which is common when functions accept multiple distinct types. #2279
• Default to "m" ABI tag when choosing libpython link name for CPython 3.7 on Unix. #2288
• Allow to compile "abi3" extensions without a working build host Python interpreter. #2293

Fixed

• Crates depending on PyO3 can collect code coverage via LLVM instrumentation using stable Rust. #2286
• Fix segfault when calling FFI methods PyDateTime_DATE_GET_TZINFO or PyDateTime_TIME_GET_TZINFO on datetime or time without a tzinfo. #2289
• Fix directory names starting with the letter n breaking serialization of the interpreter configuration on Windows since PyO3 0.16.3. #2299