Support for Windows on ARM wheels

PR Description:

The adoption of Windows on ARM (WoA) devices is steadily increasing, yet many Python wheels are still not available for this platform.
GitHub Actions now offer native CI runners for Windows on ARM devices (windows-11-arm), enabling automated builds and testing.
Currently, official statsmodels Python wheels are not provided for Windows ARM64 and thus users/developers were facing difficulties using popular statsmodels library natively.
Now SciPy and NumPy are released for Windows on ARM, having statsmodel wheels for Windows on ARM would be highly appreciated :)

Need a pandas release that supports arm64 on Windows in order to test, which is required before we can release.

First attempt failed due to changes in NumPy's npymath library structure.

#xref numpy/numpy#29747

@bashtage

@bashtage I have a local Windows on ARM setup, let try to reproduce from my side and let you know

main is now building on win-arm64

Probably won't be on PyPI until the next release, and we will likely need win-arm64 pandas since it is required for the tests.

@bashtage

@bashtage this is really great and highly appreciable, thanks for adding the support. I could see that Pandas already supports Windows ARM builds in the wheel builder, might be releasing wheels in the next minor release. Could you share some information (ETA) on the next official release for statsmodel?

Ideally it would be sometime this fall, but hard to say. First pandas, then maybe some patches, and only then would be put out a wheel on PYPi. This said, dailies will be available on anaconda in the coming days.

Available now at

https://anaconda.org/scientific-python-nightly-wheels/statsmodels/files

Thanks, will keep an eye on Panda's release

There are many failing tests when trying to run on Windows on Arm. I don't have any hardward so would really need someone to work through these failures and find the root cause (and submit PRs).

Small correction - the failures are actually segfaults. Most appear to be related to models that use the Poisson. I suspect that there is an issue somewhere in SciPy or NumPy as we don't have any compiled code in these models.

First hint as to the ultimate source

2025-10-30T18:09:48.5055765Z   File "C:\Users\runneradmin\AppData\Local\Temp\cibw-run-w3cdw1ru\cp313-win_arm64\venv-test\Lib\site-packages\numpy\linalg\_linalg.py", line 669 in inv
2025-10-30T18:09:48.5056878Z   File "C:\Users\runneradmin\AppData\Local\Temp\cibw-run-w3cdw1ru\cp313-win_arm64\venv-test\Lib\site-packages\statsmodels\genmod\generalized_linear_model.py", line 1387 in _fit_gradient
2025-10-30T18:09:48.5058174Z   File "C:\Users\runneradmin\AppData\Local\Temp\cibw-run-w3cdw1ru\cp313-win_arm64\venv-test\Lib\site-packages\statsmodels\genmod\generalized_linear_model.py", line 1305 in fit
2025-10-30T18:09:48.5059299Z   File "C:\Users\runneradmin\AppData\Local\Temp\cibw-run-w3cdw1ru\cp313-win_arm64\venv-test\Lib\site-packages\statsmodels\base\_penalized.py", line 213 in fit
2025-10-30T18:09:48.5060682Z   File "C:\Users\runneradmin\AppData\Local\Temp\cibw-run-w3cdw1ru\cp313-win_arm64\venv-test\Lib\site-packages\statsmodels\base\_screening.py", line 297 in screen_exog
2025-10-30T18:09:48.5061956Z   File "C:\Users\runneradmin\AppData\Local\Temp\cibw-run-w3cdw1ru\cp313-win_arm64\venv-test\Lib\site-packages\statsmodels\base\tests\test_screening.py", line 160 in test_glmpoisson_screening

Looks like a segfault in matrix inversion via NumPy.

@bashtage

@bashtage segfaults on scipy sides are gone. pandas just got 3.0 release with WoA support.

We can see what happens now then.

@bashtage

@bashtage I don't see any nighthly builds anymore for some reason... What's the current state of this effort? Any new blocking issues?

They seem to be here:

https://anaconda.org/channels/scientific-python-nightly-wheels/packages/statsmodels/files

Oh, you are looking for ARM. I don't think we've had ARM succeed yet, and it isn't part of the nightly build.

@bashtage

@bashtage I bet it was there last September. What's the current issues? Still some instabilities have occurs? Could you point me to the logs?

The test suit segfailts in np.linalg.inv. See the log here https://github.com/MacPython/statsmodels-wheels/actions/runs/23712035908/job/69073225754

I think that's a known issue: numpy/numpy#29442

After skipping two tests that segfault on win-arm64, the test suite passes. The next challenge is determinine how to get the nightlies to upload. We use miniforge which does not have an arm64 installer. Any ideas?

Two options:

Switch to building all wheels, then accumulate them all in a single job with artifact-upload/artifact-download, then upload them in a single job running on Linux x86-64 (scipy-release just switched to this, TBD if there are any issues with rate limits).
Handle win-arm64 specially like so: https://github.com/numpy/numpy-release/blob/2c5bc070bce1ac72bafdb66f115f1446916e67a2/.github/workflows/wheels.yml#L150-L168

I was wrong about the passing of Windows-ARM64. It is still consistently segfaulting inv operations, e.g.,

    File "C:\Users\runneradmin\AppData\Local\Temp\cibw-run-3ijbha8c\cp313-win_arm64\venv-test\Lib\site-packages\numpy\linalg\_linalg.py", line 648 in inv
    File "C:\Users\runneradmin\AppData\Local\Temp\cibw-run-3ijbha8c\cp313-win_arm64\venv-test\Lib\site-packages\statsmodels\tsa\statespace\kalman_smoother.py", line 1687 in smoothed_state_gain

This needs to be fixed upstream (OpenBLAS or NumPy, depending) before it will be possible to release for Windows-ARM64.

If anyone has a Windows-ARM64 devide wiht a different CPU, it would be great if you could see if the statsmodels test suit can run on your CPU (to see if it is a kernel-specific bug in OpenBLAS)

I will add the failing Windows-ARM64 build so that if it is fixed upstream, the wheels will start showing up in nightlies.

xref MacPython/statsmodels-wheels#191

I will add the failing Windows-ARM64 build so that if it is fixed upstream, the wheels will start showing up in nightlies.

xref MacPython/statsmodels-wheels#191

it shows up now https://anaconda.org/channels/scientific-python-nightly-wheels/packages/statsmodels/files?file_q=win_arm64 and also in pypi. so it is resolved ?

I think it is resolved. They still fail the test suite but this is upstream to us.

Navigation Menu

Repository navigation

Support for Windows on ARM wheels #9629

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Support for Windows on ARM wheels #9629

Description

Activity

bashtage commented on Sep 12, 2025

bashtage commented on Sep 12, 2025

MGN-GIT commented on Sep 12, 2025

bashtage commented on Sep 13, 2025

MGN-GIT commented on Sep 13, 2025

bashtage commented on Sep 13, 2025

bashtage commented on Sep 13, 2025

MGN-GIT commented on Sep 13, 2025

bashtage commented on Oct 30, 2025

bashtage commented on Oct 30, 2025

bashtage commented on Oct 31, 2025

khmyznikov commented on Jan 21, 2026

bashtage commented on Jan 21, 2026

khmyznikov commented on Mar 27, 2026

bashtage commented on Mar 28, 2026

bashtage commented on Mar 28, 2026

khmyznikov commented on Mar 28, 2026

bashtage commented on Mar 29, 2026

rgommers commented on Mar 29, 2026

bashtage commented on Mar 31, 2026

rgommers commented on Mar 31, 2026

bashtage commented on Apr 1, 2026

bashtage commented on Apr 1, 2026

bashtage commented on Apr 1, 2026

xieofxie commented on Sep 23, 2026

bashtage commented on Sep 28, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions