Lightlines
By nvidia

tilegym-monkey-patch-kernels-to-transformers

nvidia1.6k installs

Integrate and validate faster kernels for your language model.

What it does

Adapt Hugging Face transformers models to run TileGym kernels and develop additional cuTile kernels with correctness and performance checks.

When to use it

Use when the user requests integrating TileGym kernels into transformers models.

How to use it

Give it a target model ID and the project; it prepares the experiment environment, integrates kernels, and develops additional kernels after setup clarification and confirmation.

What you provide

  • target model ID
  • An existing project

Good to know

  • starts helper agents, number unstated
  • acts without asking

Latest review

What you need · 8

Experiments require a GPU supported by TileIR; the instructions name Ampere, Ada, and Blackwell.

Requires Docker with GPU access to build and run the experiment container; a non-Docker environment requires an explicit user request.


Access · 3

This skill

research literature

Read

Searches external research literature

docker

Execute

Starts GPU experiment containers