cuda-driver-dev may need to expose the driver stub to LD_LIBRARY_PATH

[leofang]

In pre-CUDA 12, it's expected that the libcuda.so stub is default loadable, but it seems this assumption no longer holds, and I have to tweak it so that the import cupy test (which loads the stub) works with CUDA 12.

[jakirkham]

Looking to add cuda-cudart-dev_{{ target_platform }} as a requirement of cuda-nvcc-impl_{{ target_platform }} in PR ( conda-forge/cuda-nvcc-feedstock#11 ) to ensure it is present and near the compiler when it links it. Hoping that is enough

Would be a little worried about adding the stub libraries to LD_LIBRARY_PATH as it can cause issues in runtime environments (where maybe a compiler and these libraries are also needed for CuPy or Numba). Namely the stub libraries get picked up by the library loader instead of the actual libraries

So would suggest we try and see if we can solve this in the compiler package (maybe tweaking the search paths of the compiler itself)

[isuruf]

In pre-CUDA 12, it's expected that the libcuda.so stub is default loadable, but it seems this assumption no longer holds

This is not true. Pre-CUDA 12, there is a /usr/local/lib/libcuda.so.1 in the docker image we use which loaded by the dynamic linker. With CUDA 12, we use the regular docker image without that library.