BUG: function built with `-F pyo3/abi3t` raises `TypeError` at runtime (CPython 3.15.0b2-free-threaded))

[neutrinoceros]

Context: I'm trying out PEP 803 support in neutrinoceros/rlic#423.
I found that function bindings generated for cp315-abi3.abi3t produced a TypeError at runtime against Python 3.15.0b2 (free-threaded) + numpy 2.6.0dev0 (nightly cp315-cp315t wheel). Here's a minimal reproducer¹

use pyo3::prelude::*;
use numpy::borrow::PyReadonlyArray2;

#[pymodule]
fn _core<'py>(_py: Python<'py>, m: &Bound<'py, PyModule>) -> PyResult<()> {
    use pyo3::prelude::*;

    #[pyfunction]
    fn func<'py>(
        _py: Python<'py>,
        _array: PyReadonlyArray2<'py, f64>,
    ) -> i64 {
        0
    }
    m.add_function(wrap_pyfunction!(func, m)?)?;

    Ok(())
}

>>> import reprod
>>> import numpy as np
>>> reprod.func(np.eye(2, dtype="float64"))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
    reprod.func(np.eye(2, dtype="float64"))
    ~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: 'ndarray' object is not an instance of 'ndarray'

adjacent to #547

Footnotes

1. minimal, but not self-contained. I can share a more fleshed out example if it helps. I just assumed the rest wasn't important. ↩

[Icxolu]

I can reproduce this problem and managed to observe that this dimensionality check failed:

rust-numpy/src/array.rs

Lines 163 to 169 in 4ef0102

	// Check if the dimensionality matches `D`.
	let src_ndim = array.ndim();
	if let Some(dst_ndim) = D::NDIM {
	if src_ndim != dst_ndim {
	return Err(DimensionalityError::new(src_ndim, dst_ndim).into());
	}
	}

Python 3.15.0b2 free-threading build (main, Jun 11 2026, 04:25:23) [MSC v.1951 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> from pyo3_scratch.whatever import func
>>> import numpy as np
>>> e = np.eye(2, dtype="float64")
>>> func(e)
[..\pyo3-workspace\rust-numpy\src\array.rs:153:16] check(ob.py(), ob.as_ptr()) = 1
[..\pyo3-workspace\rust-numpy\src\array.rs:164:24] array.ndim() = 0
[..\pyo3-workspace\rust-numpy\src\array.rs:165:33] D::NDIM = Some(
    2,
)
Traceback (most recent call last):
  File "<python-input-3>", line 1, in <module>
    func(e)
    ~~~~^^^
TypeError: 'ndarray' object is not an instance of 'ndarray'
while processing '_array'

In the version specific build it succeeds as expected:

Python 3.15.0b2 free-threading build (main, Jun 11 2026, 04:25:23) [MSC v.1951 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> from pyo3_scratch.whatever import func
>>> import numpy as np
>>> e = np.eye(2, dtype="float64")
>>> func(e)
[..\pyo3-workspace\rust-numpy\src\array.rs:153:16] check(ob.py(), ob.as_ptr()) = 1
[..\pyo3-workspace\rust-numpy\src\array.rs:164:24] array.ndim() = 2
[..\pyo3-workspace\rust-numpy\src\array.rs:165:33] D::NDIM = Some(
    2,
)
[..\pyo3-workspace\rust-numpy\src\array.rs:174:12] !src_dtype.is_equiv_to(&dst_dtype) = false
0

My best guess currently is that there are ABI changes in npyffi (most likely in object.rs) for ABI3T. There was no specific work yet to support abi3t, so I would not be surprised if something is missing.

cc @ngoldbaum

[ngoldbaum]

@kumaraditya303 updated the headers for abi3t support: numpy/numpy#31091

The FFI bindings over here are going to need to be updated to match.

[ngoldbaum]

(maybe npyffi could grow something like pyo3's ffi-check macro?)

[Icxolu]

Thanks @ngoldbaum. I went though this and opened #556 with the initial changes to make this work. GIL enabled CI is still green and locally the failing abi3t example from above works now. We still would need to figure out CI testing for this though.

@neutrinoceros Maybe you want to try building against #556 and see if that would work for your example?

[neutrinoceros]

Thanks a lot ! I'll try it out in my CI first thing in the morning !