NVIDIA Agentic AI Translates CUDA Tile Kernels From Python to Rust

NVIDIA has released an AI agent skill in the TileGym repository that translates cuTile Python and Triton-TileIR kernels into cuTile Rust. The skill was used to port all 24 public TileGym operators, covering roughly 40 GPU kernels, including element-wise operations, flash-attention decode, MLA, and MoE models. According to NVIDIA, the translated operators achieve 99.5% of cuTile Python performance on average, and benchmarking on an NVIDIA DGX B200 shows every operator exceeds a 0.95 geometric-mean speedup versus cuTile Python.

The translation follows a bounded multi-agent pipeline with machine-checkable verdicts at each stage. Structural equivalence is verified by diffing the generated IR against reference Tile IR before functional testing. The pipeline validates each stage with scripts and Tile IR diffs before proceeding. Semantic checks in the Python wrapper and ABI checks in the C-ABI layer guard against silent corruption.

What the Rust port changes

cuTile Rust extends Rust ownership to tile-based GPU kernels, splitting mutable outputs into disjoint pieces. This makes explicit type and shape constraints visible to rustc, enabling compile-time checks. One key difference: cuTile Python JIT specializes kernels implicitly at call time, while cuTile Rust requires explicit specialization in the kernel signature. Rust kernels embed the kernel AST, and the first launch specializes const-generic values and caches the compiled cubin.

The C-ABI layer exports one symbol per operator, passing tensor descriptors without copying or allocating. Tensor descriptors consist of pointer, rank, shape array, and strides array, mirrored between Rust and Python. Supported dtypes in the Rust ABI include f32, f16, bf16, i32, i64, f8e5m2, and f8e4m3fn.

When a Python kernel has implicit constant branches, the Rust translation creates multiple structural entries. Setting the TileGym backend to "cutile-rs" routes the same operator API to the Rust kernels. Converted kernels can also be used directly from Rust applications via the cutile crate, bypassing the FFI.

The agent skill is shipped in the TileGym repository for users to translate their own kernels. For inference engineers, this means a path to Rust-based GPU kernels with compile-time safety checks, without manually rewriting each operation. The performance parity reported by NVIDIA suggests the ownership model does not introduce overhead in practice, though independent benchmarks would be needed to confirm.