The current implementation of cuda.cudart.cudaRuntimeGetVersion() hard-codes the runtime version, rather than querying the runtime for its version. This results in incorrect runtime versions if the runtime version is different from the version of cuda-python.

Additional context

A workaround used in rapidsai/rmm#946 is to use numba's API for this instead:

import numba.cuda

def cudaRuntimeGetVersion():
    major, minor = numba.cuda.runtime.get_version()
    return major * 1000 + minor * 10