Conflict with NumPy in test/test_stltypes.py #98

Activity

1. aaronj0 commented on Sep 15, 2026

   

   aaronj0

   on Sep 15, 2026

   Collaborator

   More actions

   Hi, thanks for the report.

   I cannot reproduce this so far on linux and the CI currently tests numpy 2.5.3. Note that test_stltypes.py already imports numpy in the corresponding tests (in favour of globally importing it):

   cppjit/test/test_stltypes.py

   Lines 690 to 698 in a14171f

   	def test18_array_interface(self):
   	"""Test usage of __array__ from numpy"""
   	import cppjit
   	try:
   	import numpy as np
   	except ImportError:
   	skip("numpy is not installed")

   One possibility is that two libstdc++ copies are reachable in your environment: importing cppjit first pins the older one for the whole process, and the global import numpy loads the newer libstdc++ that numpy was built against before cppjit can pin the old one. You can test this by now adding an import cppjit before import numpy (after import sys) which should retrigger the previous error messages.

   Can you share a snippet of the error messages and your environment config: OS and architecture, Python version, numpy version, and cppjit installation mode (source/pip)? With that info this should be quick to pin down.
2. chofchof commented on Sep 15, 2026

   

   chofchof

   on Sep 15, 2026

   Author

   More actions

   @aaronj0, thanks for a detailed answer.

   One possibility is that two libstdc++ copies are reachable in your environment: importing cppjit first pins the older one for the whole process, and the global import numpy loads the newer libstdc++ that numpy was built against before cppjit can pin the old one. You can test this by now adding an import cppjit before import numpy (after import sys) which should retrigger the previous error messages.

   You were right. I do have two copies of libstdc++.so.6 - one from the OS (AlmaLinux 10) at /usr/lib64/libstdc++.so.6, and another from Conda $CONDA_PREFIX/lib/libstdc++.so.6. After adding import cppjit before import numpy, I was able to see the same error messages again.

   Can you share a snippet of the error messages and your environment config: OS and architecture, Python version, numpy version, and cppjit installation mode (source/pip)? With that info this should be quick to pin down.

   Sure. I installed cppjit following the "Development build (pip editable)" instructions from this GitHub page. Below are the details of my environment.

   $ uname -a
   Linux alma59 6.12.0-211.53.1.el10_2.x86_64 #1 SMP PREEMPT_DYNAMIC Fri Sep 11 07:00:06 EDT 2026 x86_64 GNU/Linux
   $ cat /etc/almalinux-release
   AlmaLinux release 10.2 (Lavender Lion)
   $ dnf grouplist installed
   ... (skip) ...
   Installed Groups:
      Container Management
      Development Tools
      Headless Management
   $ python --version
   Python 3.14.7
   $ which python
   ~/miniforge3/envs/llvm22/bin/python
   $ python -c 'import numpy; print(numpy.__version__)'
   2.5.3
   

   error_message.log
3. aaronj0

   linked a pull request that will close this issueImprove failure, fix crashes in conda environments/older libstdc++ #112on Sep 26, 2026
4. aaronj0

   self-assigned this

   on Oct 6, 2026