You are not logged in.
Hi,
when upgrading this morning, numpy moved from 2.3.5-1 to 2.4.0-1, and now when I try to import it I got:
$ python
Python 3.13.11 (main, Dec 7 2025, 13:01:45) [GCC 15.2.1 20251112] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy as np
Illegal instruction (core dumped) python I reckon that my cpu is rather old : Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz, but it is still sufficient for it's duty.
Already tested:
- Downgrade to 2.3.5-2 -> ModuleNotFoundError: No module named 'numpy'
- Downgrade to 2.3.5-1 -> Works.
Any advice on what to do to be ready for Python 3.14 ?
Thanks
Offline
Hello. It’s a bug in packaging
Numpy 2.4 now uses x86_64-v2 as the minimum architecture with optimizations enabled, and enables that by default, silently even when the platform doesn’t support it. Which generates invalid instructions. This one to be precise:
pinsrq $0x1,0x8b499b(%rip),%xmm0The announcement is hidden deep in release notes, as a single sentence inside another section.
The workaround, until Arch adjusts the package, is to rebuilt the package with a changed PKGBUILD:
python -m build --wheel --no-isolation \
-Csetup-args="-Dblas=cblas" \
-Csetup-args="-Dlapack=lapack" \
-Csetup-args="-Denable-openmp=true" \
-Csetup-args="-Dcpu-baseline=none"To align with Google Highway, enabling v1 optimizations on v1 platforms is also blocked.
Last edited by mpan (2026-01-06 13:11:36)
Paperclips in avatars? | Sometimes I seem a bit harsh — don’t get offended too easily!
Offline
Thanks, I'm relieved to ear it will be addressed.
I'll live with 2.3.5 until then.
Offline
I was about to write to arojas about this, but I discovered it’s fixed (as of 2.4.0-3).
Feel free to update. You should: python 3.14 hit the repos and old python-* packages aren’t going to work.
Paperclips in avatars? | Sometimes I seem a bit harsh — don’t get offended too easily!
Offline