TileGym

Why we included this project

TileGym is a practical place to learn tile-based GPU programming with CUDA Tile, built around runnable examples rather than slides. It ships working kernels for common deep learning operators, micro-benchmarks to compare their efficiency, and end-to-end transformer examples that drop those kernels into real models like Llama 3.1 and DeepSeek V2. That makes it a solid learning path for developers new to tile programming and a handy reference for teams tuning their own LLM inference kernels. The same operators also appear across several backends, including CUDA Tile C++, Triton's tile IR, and experimental Julia and Rust variants, so you can see how one idea maps to different toolchains. Just keep in mind it targets recent NVIDIA hardware and CUDA 13.1+, so treat it as a hands-on kernel workshop rather than a drop-in runtime for older GPUs.