cuTile Rust Brings Ownership-Safe GPU Kernels via JIT Compilation

cutile-rs extends Rust's borrow checker across GPU launch boundaries by partitioning mutable tensors into disjoint pieces and sharing immutable ones, preventing data races at compile time. Kernels are JIT-compiled from a captured Rust AST through CUDA Tile IR to cubin, supporting sync, async, and CUDA graph replay. Early-stage research project with expected API instability.