Add pre-compiled wheel for Python 3.13

[dperdios]

It seems like wheels are pre-compiled for all supported versions of Python except for 3.13 (https://github.com/Forest-Neurotech/mach/releases/tag/v0.1.0).

It seems like 313 is targeted for automatic build via CI. Unsure why it failed to build that wheel.

mach/pyproject.toml

Lines 229 to 236 in c133065

	[tool.cibuildwheel]
	build = [
	"cp39-manylinux_x86_64",
	"cp310-manylinux_x86_64",
	"cp311-manylinux_x86_64",
	"cp312-manylinux_x86_64",
	"cp313-manylinux_x86_64",
	]

Would be cool to include a wheel for 3.13 too :)

[dperdios]

Just did a quick test in a fresh python 3.13 environment

(mach-beamform-py313) dimitris@surveyor:~/dev/mach-beamform-py313$ uv pip list
Package          Version
---------------- -------
array-api-compat 1.12.0
cupy-cuda12x     13.6.0
einops           0.8.1
fastrlock        0.8.3
jaxtyping        0.3.3
mach-beamform    0.1.0
numpy            2.3.5
wadler-lindig    0.1.7
(mach-beamform-py313) dimitris@surveyor:~/dev/mach-beamform-py313$ python
Python 3.13.6 (main, Aug  8 2025, 17:08:22) [Clang 20.1.4 ] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import mach
Traceback (most recent call last):
  File "<python-input-0>", line 1, in <module>
    import mach
  File "/home/dimitris/dev/mach-beamform-py313/.venv/lib/python3.13/site-packages/mach/__init__.py", line 17, in <module>
    from . import geometry, kernel, wavefront
  File "/home/dimitris/dev/mach-beamform-py313/.venv/lib/python3.13/site-packages/mach/kernel.py", line 12, in <module>
    from ._cuda_impl import (
    ...<2 lines>...
    )
ModuleNotFoundError: No module named 'mach._cuda_impl'
>>> 

For comparison in a fresh python 3.12

(mach-beamform-py312) dimitris@surveyor:~/dev/mach-beamform-py312$ uv pip list
Package          Version
---------------- -------
array-api-compat 1.12.0
cupy-cuda12x     13.6.0
einops           0.8.1
fastrlock        0.8.3
jaxtyping        0.3.3
mach-beamform    0.1.0
numpy            2.3.5
wadler-lindig    0.1.7
(mach-beamform-py312) dimitris@surveyor:~/dev/mach-beamform-py312$ python
Python 3.12.3 (main, Nov  6 2025, 13:44:16) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import mach
>>> 

[charlesbmi]

Thanks for the report!

The Python 3.13 wheel was getting built as part of the STABLE_ABI,

but the nanobind config had a bug - it was missing a CMake package that caused it to just put the 3.12 .so in the *-cp312-abi3-*.whl, so it was missing the .so in Python 3.13.

I read that nanobind STABLE_ABI can have some very slight performance issues, so I went ahead and just made it build a separate wheel for Python 3.13: #52

I triggered a release-build, feel free to check in >10 minutes if the PyPI/Github release: https://github.com/Forest-Neurotech/mach/releases/tag/v0.1.1

I'll test sometime later in the week myself too.

[charlesbmi]

Fixed in https://github.com/Forest-Neurotech/mach/releases/tag/v0.1.1 , closing

Example on Linux (with no GPU, so we appropriately a driver error)

$ uvx --with mach-beamform --with numpy python -c "import mach; print(mach.__version__)"
Installed 6 packages in 16ms
<frozen importlib._bootstrap>:488: RuntimeWarning: [mach] CUDA driver version (0.0) is too old for code compiled with NVCC 12.3.107. You may see kernel launch failures or crashes.
→ Please update your NVIDIA driver to version 12.3 or newer.
0.1.1

Env info for reference:

$ lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description:    Ubuntu 24.04.3 LTS
Release:        24.04
Codename:       noble

Note: there is separately an implicit dependency for numpy through jaxtyping, worked-around on #53 , filed patrick-kidger/jaxtyping#360

[dperdios]

Thanks - works perfectly! The output on a CUDA-capable computer :)

$ uvx --with mach-beamform --with numpy python -c "import mach; print(mach.__version__)"
Installed 6 packages in 10ms
0.1.1

Good you opened an issue for the implicit numpy dep (forgot to open one about that yesterday)