NVIDIAOfficial·7 Agent Skills Included

TileGym

Build, convert, and optimize CUDA Tile GPU kernels for LLMs

Writes, ports, and tunes CUDA Tile GPU kernels for deep learning operators like attention, matmul, and normalization. Converts kernels between cuTile Python, Triton, Rust, and Julia, and patches them into Hugging Face transformer models. Removes guesswork from GPU programming with step-by-step checklists, autotuning templates, and benchmark-driven optimization loops.

807|85

↗

npx skills add NVIDIA/TileGym --all -g -y

All Skills in This Repository (7)

Pure Emerald Level Indicators

📦 In Repo

NVIDIA

converting-cutile-to-triton

Translate cuTile GPU kernels to Triton with dual-layout and META-driven autotuning.

monkey-patch-kernels-to-transformers

Monkey-patch TileGym CUDA kernels into HuggingFace Transformers models.

cutile-python

Guide cuTile kernel development from planning to validation against a PyTorch reference.

converting-cutile-to-julia

Convert cuTile Python kernels into cuTile.jl implementations with 1-based indexing and broadcasting.

adding-cutile-kernel

Automate adding cuTile kernel operators to TileGym with registration, backend, tests, and benchmarks.

improve-cutile-kernel-perf

Profile cuTile kernels and apply one optimization per iteration.

tilegym-converting-cutile-triton-to-cutile-rs

Convert Triton-TileIR and cuTile-Python GPU kernels to cutile-rs Rust via a multi-agent pipeline.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install TileGym?▼

Run `npx skills add NVIDIA/TileGym --all -g -y` in your terminal to install all skills in this suite globally.

What is TileGym used for?▼

TileGym is a CUDA Tile kernel library with tutorials and examples for tile-based GPU programming, plus ready kernels for LLMs like Llama 3.1 and DeepSeek V2.

Can AI convert cuTile kernels to Triton or Rust?▼

Yes. Dedicated skills guide your agent through porting cuTile kernels to Triton, Julia, or cutile-rs with correctness tests and performance gates.

How do I speed up a slow GPU kernel?▼

The autotuning and performance skills run an iterative loop that profiles the kernel, applies one optimization at a time, and keeps only changes that measurably reduce latency.

What hardware does TileGym require?▼

You need CUDA 13.1 or newer and an NVIDIA Blackwell GPU such as B200 or RTX 5090; Ampere GPUs like A100 work with CUDA 13.2+.