Pin llvm-openmp=11 when building wheels for macOS

[jtilly]

Closes #628.

I was able to reproduce the issue using the intel-based macOS runner: I'm building the wheel using cibuildwheel and then I'm installing the wheel and running the code snippet from the issue:

jobs:
  macos:
    name: "issue-628"
    runs-on: macos-11
    steps:
      - uses: actions/checkout@v3
      - name: Build wheels
        uses: pypa/cibuildwheel@v2.13.0
        env:
          CIBW_ARCHS_MACOS: x86_64
      - uses: actions/setup-python@v4
        with:
          python-version: '3.9.15'
      - run: |
          pip install ./wheelhouse/*.whl
          python .github/workflows/issue_628.py

This produces:

/Users/runner/work/_temp/d0cacb75-060f-4cd9-9f66-b7d751687330.sh: line 4:  6610 Segmentation fault: 11  python .github/workflows/issue_628.py

When I pin llvm-openmp to version 11 (like scikit-learn), everything runs through (code, output).

[MarcAntoineSchmidtQC]

Amazing! Thanks Jan.

[MarcAntoineSchmidtQC]

Let's wait that the CI finishes but all good from my side.

[MarcAntoineSchmidtQC]

Let's wait that the CI finishes but all is good from my side. Not sure how we can make sure this doesn't happen. The wheels are getting tested on the wheel-building architecture, but the problem happens when you install this on a different machine.

[jtilly]

The wheels are getting tested on the wheel-building architecture, but the problem happens when you install this on a different machine.

I was able to reproduce the issue on the same machine. See https://github.com/Quantco/glum/actions/runs/5145139930/jobs/9262364272#step:5:63. We could add something along these lines to our CI pipeline. Build the wheels, install the wheels, run some code.

[jtilly]

Possibly related?

• BLD Fixes osx build by downgrading to 11.X [cd build] scikit-learn/scikit-learn#21227 (comment)
• Segmentation fault on MacOS with libomp 12.0.0 dmlc/xgboost#7039
• CI Use conda-forge's openmp in osx to build wheels scikit-learn/scikit-learn#22109 (review)