Lightlines
By nvidia

tilegym-adding-cutile-kernel

nvidia1.9k installs

Add a cuTile operator with tests and a benchmark.

What it does

Adds a GPU operator to TileGym with cuTile implementation, dispatch registration, exports, tests, and a benchmark.

When to use it

Use when adding or implementing a cuTile operator in TileGym, or asking how to register one.

How to use it

Give it an operator to add to a TileGym project; it implements and registers the cuTile backend, adds tests and a benchmark, and runs verification.

What you provide

  • An existing project

Latest review

What you need · 6

The backend and tests import torch for tensor creation and the reference implementation.

The backend implementation imports cuda.tile and uses its kernel decorator.

The operator targets a GPU, and its test creates tensors on the CUDA device.

The required benchmark imports triton and triton.testing for benchmark configuration and timing.


Access · 0

No additional actions listed.