Skip to content


Repository files navigation


Actions Status Crate Minimum rustc 1.48 Documentation codecov

Rust bindings for the NumPy C-API.

API documentation


  • Rust >= 1.48.0
    • Basically, our MSRV follows the one of PyO3
  • Python >= 3.7
    • Python 3.6 support was dropped from 0.16
  • Some Rust libraries
  • numpy installed in your Python environments (e.g., via pip install numpy)
    • We recommend numpy >= 1.16.0, though older versions may work


Write a Python module in Rust

Please see the simple example for how to get started.

There are also examples using ndarray-linalg and rayon.

name = "rust_ext"
crate-type = ["cdylib"]

pyo3 = { version = "0.21", features = ["extension-module"] }
numpy = "0.21"
use numpy::ndarray::{ArrayD, ArrayViewD, ArrayViewMutD};
use numpy::{IntoPyArray, PyArrayDyn, PyReadonlyArrayDyn, PyArrayMethods};
use pyo3::{pymodule, types::PyModule, PyResult, Python, Bound};

fn rust_ext<'py>(_py: Python<'py>, m: &Bound<'py, PyModule>) -> PyResult<()> {
    // example using immutable borrows producing a new array
    fn axpy(a: f64, x: ArrayViewD<'_, f64>, y: ArrayViewD<'_, f64>) -> ArrayD<f64> {
        a * &x + &y

    // example using a mutable borrow to modify an array in-place
    fn mult(a: f64, mut x: ArrayViewMutD<'_, f64>) {
        x *= a;

    // wrapper of `axpy`
    #[pyo3(name = "axpy")]
    fn axpy_py<'py>(
        py: Python<'py>,
        a: f64,
        x: PyReadonlyArrayDyn<'py, f64>,
        y: PyReadonlyArrayDyn<'py, f64>,
    ) -> Bound<'py, PyArrayDyn<f64>> {
        let x = x.as_array();
        let y = y.as_array();
        let z = axpy(a, x, y);

    // wrapper of `mult`
    #[pyo3(name = "mult")]
    fn mult_py<'py>(a: f64, x: &Bound<'py, PyArrayDyn<f64>>) {
        let x = unsafe { x.as_array_mut() };
        mult(a, x);


Execute a Python program from Rust and get results

name = "numpy-test"

pyo3 = { version = "0.21", features = ["auto-initialize"] }
numpy = "0.21"
use numpy::{PyArray1, PyArrayMethods};
use pyo3::{types::{IntoPyDict, PyAnyMethods}, PyResult, Python};

fn main() -> PyResult<()> {
    Python::with_gil(|py| {
        let np = py.import_bound("numpy")?;
        let locals = [("np", np)].into_py_dict_bound(py);

        let pyarray = py
            .eval_bound("np.absolute(np.array([-1, -2, -3], dtype='int32'))", Some(&locals), None)?

        let readonly = pyarray.readonly();
        let slice = readonly.as_slice()?;
        assert_eq!(slice, &[1, 2, 3]);


Dependency on ndarray

This crate uses types from ndarray in its public API. ndarray is re-exported in the crate root so that you do not need to specify it as a direct dependency.

Furthermore, this crate is compatible with multiple versions of ndarray and therefore depends on a range of semver-incompatible versions, currently >= 0.13, < 0.16. Cargo does not automatically choose a single version of ndarray by itself if you depend directly or indirectly on anything but that exact range. It can therefore be necessary to manually unify these dependencies.

For example, if you specify the following dependencies

numpy = "0.21"
ndarray = "0.13"

this will currently depend on both version 0.13.1 and 0.15.3 of ndarray by default even though 0.13.1 is within the range >= 0.13, < 0.16. To fix this, you can run

cargo update --package ndarray:0.15.3 --precise 0.13.1

to achieve a single dependency on version 0.13.1 of ndarray.


We welcome issues and pull requests.

PyO3's is a nice guide for starting.

Also, we have a Gitter channel for communicating.