openai-whisper: use llvm@20 (pre-release testing)

[cho-m]

Testing pre-releases of llvmlite and numba

CMake addition is for llvmlite. Same for environment variable changes:

• https://github.com/numba/llvmlite/blob/main/docs/source/admin-guide/install.rst#compiling-llvmlite

RC tracking issues:

• llvmlite 0.45.0-rc1 Checklist numba/llvmlite#1254
• Numba 0.62.0-rc1 Checklist numba/numba#10170

---

Tested:

• system libexec/"bin/python", "-m", "llvmlite.tests"
• system libexec/"bin/python", "-m", "numba.runtests", "-m"

[esc]

@cho-m looks like the test failure was some sort of glitch that caused a timeout?

Context: Numba release manager here, I am checking on dependents and the stability of the release candidate.

[cho-m]

@cho-m looks like the test failure was some sort of glitch that caused a timeout?

Yeah, you can ignore the failure as it is unrelated to numba/llvmlite. Likely a slow CI runner.

[cho-m]

To remove on a future run. Just checking our shared LLVM 20 works with llvmlite via test suite:

Suggested change

	# TODO: remove
	system libexec/"bin/python", "-m", "llvmlite.tests"
	system libexec/"bin/python", "-m", "numba.runtests", "-m"

[esc]

FWIW: using a shared LLVM is somewhat discouraged and this setup has been removed in conda-forge: conda-forge/llvmlite-feedstock#100 -- in case you end up with user issues and weird segfaults, this may be something to consider.

[cho-m]

Closing until stable releases are available