Make a compiler that takes Rust and compiles into some IR, then another compiler that compiles that IR into GPU machine code. Then it can work and that's going

Make a compiler that takes Rust and compiles into some IR, then another compiler that compiles that IR into GPU machine code. Then it can work and that's going to be your API (what you developed in Rust). That's the whole point of what's missing. Not some wrapper around CUDA.