sablin39 / tilelang-cuda-skills

What is tilelang-cuda-skills?

This is a curated set of AI agent skills designed to help developers write, debug, profile, and optimize GPU kernels using TileLang and raw CUDA. Think of it as a cheat sheet layer on top of AI coding assistants -- each skill defines a prompt and workflow so Claude (or similar tools) can assist with GPU development tasks more reliably. It includes skills for writing kernels from scratch, diagnosing compilation failures, measuring performance with ncu and roofline analysis, and tuning for Blackwell GPUs with CUDA 13.1 and PyTorch 2.11.

Why is it gaining traction?

The hook is workflow coherence. Rather than asking an AI to "help with my CUDA kernel" and getting generic output, these skills enforce a structured progression: write first, debug second, profile third, optimize last. For developers building differentiable operators in ML frameworks, it also covers forward-backward kernel testing with mixed precision. The CUDA skill bundles PTX ISA 9.1 and CUDA Runtime/Driver API documentation as searchable markdown files, which is useful when debugging without an IDE open.

Who should use this?

ML framework authors or researchers writing custom GPU kernels in TileLang who want AI-assisted development without reinventing the same debugging and profiling workflow repeatedly. Backend engineers working on CUDA kernels that need quick access to PTX ISA or CUDA API references during a debug session. Not for general Python developers -- the audience is narrow and the skills assume you already understand GPU execution models.

Verdict

At 19 stars with only a README and no test suite, this is an early-stage, niche resource. The credibility score of 0.9% reflects that. If you are actively working in TileLang or deep CUDA optimization and using Claude Code, these skills will save you time. For everyone else, watch the repo but wait for more community validation before integrating it into a production workflow.