That's a problem on a different level of the CUDA stack.
Having a compiler that takes a special C++ or python dialect and compiles it to GPU suitable llvm-ir and then to a GPU binary is one thing (and there's progress on that side: triton, numba, soonish mojo), being able to launch that binary without going through the nvidia driver is another problem.
Vulkan Compute's semantics are limited by SPIR-V and thus cannot implement all of the features CUDA provides (ex. there is no proper notion of a "pointer")
Also it's much more convenient to use plain C++ rather than a custom shading language, especially if you're writing complex numerical code or need some heavy templated abstractions to do powerful stuff. And the CUDA tooling itself is just much easier to use compared to Vulkan, with its seamless integration of host / device code.
You still require a Vulkan driver to do anything with it. Until last year, Nvidia hardware required a proprietary Vulkan driver (prior to Nvvk), and anything pre-Pascal still requires that.
Having a compiler that takes a special C++ or python dialect and compiles it to GPU suitable llvm-ir and then to a GPU binary is one thing (and there's progress on that side: triton, numba, soonish mojo), being able to launch that binary without going through the nvidia driver is another problem.