tilegym-adding-cutile-kernel

This skill is an NVIDIA end-to-end workflow for adding a new cuTile GPU kernel operator to the TileGym library. It solves the problem of correctly wiring a new operator through every layer of TileGym so it is dispatchable, implemented, exported, tested, and benchmarked, avoiding the common mistake of forgetting a registration step that leaves the backend implementation unloaded. It enforces a strict ordered checklist via TodoWrite, requiring each step to result in a file write or an explicit skip decision with no silent omissions.

The workflow has six steps: registering a dispatch interface in ops.py as the single entry point (whose body only raises NotImplementedError), implementing the cuTile backend kernel in a dedicated module using cuda.tile with register_impl, registering the new operator in the cutile __init__.py (imports and __all__), adding tests that import from tilegym.ops rather than the backend module and compare against a PyTorch reference across shapes and dtypes, adding a benchmark under tests/benchmark, and finally verifying with pytest and lint. It provides code templates for each step and points to existing operators like silu_and_mul and softmax as references. The bundled benchmark and skill card report NVSkills-Eval results with a 100% security score, no required credentials, and dual CC-BY-4.0/Apache-2.0 licensing.

It targets GPU kernel developers and ML systems engineers contributing operators to TileGym. Typical use cases include adding, creating, or implementing a new cuTile operator and learning how to register it properly. The output is Python source files and shell commands run against a local TileGym checkout.