RuntimeError: Cannot patch Numba: version 0.57.1 should not be patched

Question

I am trying to run rapidsAI on WSL Windows to run Machine Learning models on GPU and not CPU, i installed conda on WSL, set an env with python version 3.8.15, cuml=22.02, cudatoolkit=11.2

i try to import cuml and i get this error "RuntimeError: Cannot patch Numba: version 0.57.1 should not be patched. Set the environment variable NUMBA_CUDA_ENABLE_MINOR_VERSION_COMPATIBILITY=1 instead. See https://numba.readthedocs.io/en/stable/cuda/minor_version_compatibility.html for more details."

i tried export NUMBA_CUDA_ENABLE_MINOR_VERSION_COMPATIBILITY=1 but the error persists

Also (https://i.stack.imgur.com/Vj4fq.png)

any help please

i tried the manual here https://numba.readthedocs.io/en/stable/cuda/minor_version_compatibility.html

but still the same problem.

Answer 1

On top of what Supra requested, is there a reason why you're running RAPIDS 22.02? It's on 23.10. I'd suggest that you either do 2 things:

1. As Numba 0.57.1 was released June 2023, it seems that the conda recipe is pulling a far newer - and incompatible - numba than for what 22.02 was built. If you really need 22.02, please use our 22.02 docker containers. Yes, you can also pin numba, but other packages may also grab a later, more incompatible versions.
2. If you want to use conda to install RAPIDS, please update your RAPIDS version to our latest stable (currently 23.10).

I'll let the team know about this issue for future versions, however, as v22.02 is a year and a half old and highly deprecated, it most likely won't get fixed.