A scheduled lock file update caused the macOS arm64 CI of scikit-learn to use blis instead of the usual OpenBLAS and uncovered at least one bug in blis on that platform as a result.
I asked an agent to isolate a minimal reproducer in Python and C. It boils down to a single faulty DGEMM call. Here is a PR created by the agent that uses the macOS arm64 runner of GitHub Actions to reproduce the problem:
ogrisel/cursor-sandbox#10
All the info to reproduce the corruption are available in the linked PR but I can extract some of it here if you prefer.
A scheduled lock file update caused the macOS arm64 CI of scikit-learn to use blis instead of the usual OpenBLAS and uncovered at least one bug in blis on that platform as a result.
I asked an agent to isolate a minimal reproducer in Python and C. It boils down to a single faulty DGEMM call. Here is a PR created by the agent that uses the macOS arm64 runner of GitHub Actions to reproduce the problem:
ogrisel/cursor-sandbox#10
All the info to reproduce the corruption are available in the linked PR but I can extract some of it here if you prefer.