triton-cuda-patterns

To write a Triton CUDA kernel for elementwise, reduction, or matmul operations, use standardized programming patterns that provide ready-to-use templates to accelerate implementing core operators with minimal boilerplate.

The best way to prototype custom operators in Triton is to instantiate standardized kernel templates for elementwise, reduction, or matmul modes and adapt the block sizes for your target operator.

Handle host integration and block configuration for Triton CUDA kernels by following built-in guides that help optimize performance on CUDA devices through proper block and grid setup.

Yes, you can build a new operator by selecting a standardized Triton elementwise, reduction, or matmul pattern and adapting the block sizes to fit your specific CUDA workflow requirements.

Yes, the triton-cuda-patterns skill requires YAML frontmatter containing name, description, category, and version, followed by a Markdown body detailing the three programming modes and optional resource directories.