Tasks

[urob]

• Add sample pyproject.toml with platform-specific indexes
• Disable expensive workflows in forks
• Configure fine-grained access token
• Explore feasibility/gain of link time optimization (-Db_lto) -> Not the bottleneck for blas/lapack operations. Might yield small boost for C++ heavy code like HiGHS. But build time costs are high.
• Explore stripping test suites from wheels
• Explore adding fft, random and u-math modules (check for regressions, Intel mutes tests!) -> Not permitted by Numpy terms of usage. Also, easy to use or patch by importing those packages separately or running the bundled patch scripts.
• Don't hardcode delvewheel --exclude, fetch libraries using pkg-config or dlltree instead?
• Make openmp.pc patch a proper patch (needs patch since it's not in a repo)
• Upgrade rtools/gcc to 4.5? stale: we dropped the pipeline completetely
• Switch from mkl-sdl to mkl-dynamic-ilp64-iomp? (requires loading dlls in mkl-service)
• Clean up workflow files to remove reminiscent of native linux workflows
• Pin Python version for tools
• Pin uv-version in pyproject.toml
• Force managed-only for all uv run calls in pyproject.toml?
• Figure out how to configure dependabot to bump uv version (pip ecosystem?)
• Figure out how to bump indirect dependencies in uv.lock
• Inline script dependencies, move common components to _utils.py?
• Make releases immutable
• Make sure later releases take precedence for exactly matching asset names

[naegelejd]

This is fantastic. Thank you for maintaining this.