NVIDIA/cutile-python

cuTile Python is a programming language for writing parallel kernels on NVIDIA GPUs.

The tool addresses the challenge of GPU kernel development by providing a high-level programming model built on Tile IR, which abstracts away low-level GPU programming details. Developers write kernels in Python that are compiled to GPU code, allowing them to express parallel computation in terms of tiles—logical blocks of data—rather than managing individual threads directly. The approach requires NVIDIA Driver r580 or later and the tileiras compiler, which currently supports Blackwell and Ampere/Ada GPUs, with Hopper support planned for future versions.

The tool suits developers working on NVIDIA GPUs who want to write parallel kernels without dropping to lower-level languages. It is available through PyPI as the cuda-tile package for straightforward installation, or can be built from source for those needing the latest development version. Building from source requires a C++17 compiler, CMake, Python 3.10 or later, and CUDA Toolkit 13.1 or later. The project includes sample code and references TileGym for additional learning resources.

Maintainers typically respond to new issues and pull requests within a few days. Work in the issue tracker centers on bug reports, feature requests, and feedback collection.