NVlabs / cutile-rs

What is cutile-rs?

cuTile Rust provides a safe, tile-based kernel programming DSL for the Rust programming language. It features a safe host-side API for passing tensors to asynchronously executed kernel functions on NVIDIA GPUs. Developers get an ergonomic way to author GPU kernels directly in Rust, compiling to CUDA Tile via MLIR without raw CUDA hazards.

Why is it gaining traction?

Unlike GitHub cuTile Python bindings, NVIDIA cuTile Rust offers native Rust safety with async tensor handling and automatic partitioning for parallel tile execution. The DSL simplifies tile loads/stores while the API integrates seamlessly with Rust async runtimes for efficient kernel launches. Early adopters praise the zero-cost abstractions over writing CUDA C++ or wrestling Python FFI.

Who should use this?

Rust ML engineers building custom GPU kernels for transformers or GEMM ops on Ampere+ NVIDIA cards. HPC devs needing safe async tensor pipelines without CUDA expertise. Teams porting PyTorch/CuPy tile code to Rust for performance-critical inference.

Verdict

Promising alpha for Rust GPU hackers, but 24 stars and 1.0% credibility signal high risk—expect API breaks and setup pain (CUDA 13.2, LLVM 21, nightly Rust). Try examples if you're on Ubuntu 24.04 with sm_80+ hardware; otherwise, stick to mature alternatives like cudarc.

(187 words)