What is tilelang-cuda-skills?

This repository is a collection of "skills", structured reference guides and templates, designed to help AI coding assistants (specifically Claude Code) write, debug, profile, and optimize GPU programs. GPUs (graphics processing units) are the chips used for AI training and other parallel computations, programming them directly requires writing specialized low-level code called "kernels." The collection covers two areas. The first is TileLang, a newer higher-level language for writing GPU kernels that is easier to work with than raw CUDA. The skills walk through: writing kernels from scratch using standard patterns, diagnosing and fixing common errors, measuring performance with benchmarking tools, and tuning for maximum speed. There is also a skill for testing kernels that need to support both forward and backward passes (a requirement when using them in machine learning training). The second area is CUDA, the lower-level programming system from NVIDIA, covering debugging with specialized tools, profiling to find bottlenecks, and understanding how the GPU's instruction set works. The intended audience is developers (or AI agents) who are building high-performance GPU code and need quick, accurate reference material organized by task. The README notes the TileLang skills were originally generated and validated on specific GPU hardware with specific software versions. Rather than tutorials, these are practical checklists and pattern references meant to be loaded into an AI assistant's context during active development.