Agreed on ZLUDA being the practical choice. This project is more impressive as a "build a GPU compiler from scratch" exercise than as something you'd actually use for ML workloads. The custom instruction encoding without LLVM is genuinely cool though, even if the C subset limitation makes it a non-starter for most real CUDA codebases.
ZLUDA doesn't have full coverage though and that means only a subset of cuda codebases can be ported successfully - they've focused on 80/20 coverage for core math.
