Guide cuTile kernel development from planning to validation against a PyTorch reference.
npx skills add https://github.com/NVIDIA/TileGym --skill cutile-python
Please help me install this Agent Skill. Skill: cutile-python Source: https://github.com/NVIDIA/TileGym/tree/main/.claude/skills/cutile-python Command: npx skills add https://github.com/NVIDIA/TileGym --skill cutile-python
Expert cuTile programming guidance to write, validate, and optimize tile-based GPU kernels with structured workflows and best practices.
Run the end-to-end cuTile kernel example to generate and validate a simple kernel.
High-intent search queries and answers about installing and using this skill.
To design and validate a cuTile GPU kernel, use a structured workflow that scaffolds the kernel design, generates code using ct.load and ct.store tile APIs, and performs end-to-end validation against a PyTorch reference.
cuTile kernel development requires enforcing power-of-two tile sizes, type-safe constants, and strict grid coverage to ensure structured and error-free GPU kernel deployment.
Yes, cuTile supports orchestration-ready guidance for both single-kernel tasks and complex multi-kernel pipelines, allowing you to manage intricate GPU workflows end-to-end.
cuTile performs end-to-end validation by comparing the generated GPU kernel's output directly against a PyTorch reference, ensuring numerical correctness and expected behavior across tile operations.
PyTorch is used as the inline validation reference to verify cuTile tile APIs, providing a baseline to check correctness and ensure the generated kernel produces expected outputs.
yo-stevenAdds a cuTile GPU kernel operator to TileGm with full integration.
yo-stevenGenerate validated cuTile GPU kernels with float16 inputs and float32 accumulators.
NVIDIATranslate cuTile GPU kernels to Triton with dual-layout and META-driven autotuning.
yo-stevenConvert cuTile @ct.kernel GPU kernels to Triton @triton.jit implementations.
