Probably true, but it usually works anyway. And a lot of ML code depends on exact versions of the CUDA libraries, so you don't get a choice.