Translating CUDA Tile Operations from Python to Rust Using Agentic AI

Summary

cuTile Rust (cutile-rs) extends Rust’s ownership model to tile-based GPU kernels, and this post describes an agentic workflow for translating cuTile Python and Triton-TileIR implementations. Because all three front ends target CUDA Tile IR and share the tileiras compiler, the workflow compares reference and translated IR while making Python’s implicit specializations explicit through Rust const generics, fixed dtype symbols, and FFI checks. A bounded multi-agent pipeline covers analysis, device kernels, host and FFI code, correctness, and benchmarking, with machine-checkable verdicts controlling progress. It ported all 24 public TileGym operators, roughly 40 kernels, and reported 99.5% average cuTile Python performance across 347 paired configurations on NVIDIA DGX B200. The skill ships with TileGym and uses acceptance checks requiring correctness, backend execution, and a geomean speedup of at least 0.95.

Takeaways

1. 1

   All 24 public TileGym operators were ported, covering roughly 40 kernels from element-wise operations through flash-attention decode, MLA, and MoE models.

2. 2

   Tile IR diffs verify memory-operation families, tile shapes, reductions, cost hints, and attributes before tests, helping detect plausible translations that could fail outside test coverage.

3. 3

   Rust requires specializations to be declared in kernel signatures, so the port makes tile shapes, element types, dtype dispatch, structural branches, and ABI validation explicit.

FAQ