AI Agent Translates CUDA Code Python to Rust

An engineering team just ported 24 GPU operators, roughly 40 kernels in total, from Python to Rust without a human rewriting a single one by hand. The work relied on an AI agent built specifically to translate CUDA tile kernels from the TileGym library, which already holds a large stock of production kernels written in cuTile Python and Triton-TileIR. The target was cuTile Rust, a tile-based system for writing GPU kernels that extends Rust's ownership model to handle mutable outputs and preserve host-side ownership across kernel launches, while still letting programmers drop down to raw Tile IR when they need that control.

The kernels involved aren't simple. They span element-wise math, flash-attention decode, Multi-head Latent Attention, and mixture-of-experts models, some requiring several variants apiece. Converting them means running each one through a multi-agent pipeline that handles analysis, device kernel logic, host and FFI code, and benchmarking, with validator scripts and Tile IR diffs checking every stage before it moves forward.

The result, after all 24 operators were converted, reached 99.5% of the original Python performance on average. The harder problem sits in a language-level mismatch that the team had to solve directly.