Cannot activate SVML

$ mamba list | grep -e numba -e lib-rt
intel-cmplr-lib-rt        2025.3.1           h04fd95a_760    conda-forge
numba                     0.63.0          py314h8169c2f_0    conda-forge

$ numba -s  | grep SVML
__SVML Information__
SVML State, config.USING_SVML                 : False
SVML Library Loaded                           : True
llvmlite Using SVML Patched LLVM              : False
SVML Operational                              : False

I had this problem before when I was trying to provide context in Numba np.exp overload 6x slower than numpy · Issue #10060 · numba/numba · GitHub, but I forget how I overcome it then. Seems related to https://github.com/numba/numba/issues/9713 but there’s no obvious solution there?

It’s in a fresh environment.

Intel SVM is currently un-supported as of llvmlite 0.45.

LLVM 20 starts with llvmlite 0.45: