Use Vulkan to allocate CUDA memory that is writable from the host. See the documentation on readthedocs for further information.