# Tilegym converting cutile to triton

> Tilegym converting cutile to triton is an open-source agent skill for developer tools. Guides an agent in converting cuTile GPU kernels to Triton and debugging the result. It is available on Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf and Goose.

- **Kind:** Skill
- **Pricing:** Open source
- **Platforms:** Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, Goose
- **Website:** https://github.com/NVIDIA/skills/tree/main/skills/tilegym-converting-cutile-to-triton
- **Repository:** https://github.com/nvidia/skills
- **GitHub stars:** 4,037
- **License:** Apache-2.0
- **Skill name:** tilegym-converting-cutile-to-triton
- **Collection:** https://github.com/NVIDIA/skills
- **Categories:** developer tools
- **Listing updated:** 2026-11-18

## About

This skill walks an agent through porting cuTile kernels marked with @ct.kernel into Triton kernels marked with @triton.jit. It covers a phased workflow of analysis, conversion, validation, testing and benchmarking, with a pre-flight grep pass over the source to spot patterns that need special handling, such as transposes, casts, launches and constant arguments.

The guidance stresses tensor-descriptor loads and stores for block-shaped tiles, bracket-style Triton launches with tuple or lambda grids, and autotuning when the source used kernel configs. It also includes reference notes on API mapping, common translation pitfalls, performance regression risks, and math dtype requirements for functions like erf and exp.

For layout-dependent kernels, such as those with transpose flags, the skill directs the agent to write separate Triton kernels per layout rather than applying a transpose inside one kernel. Worked examples cover vector add, softmax, layernorm, matmul and attention, each with a cuTile source and a Triton target. Running the conversion requires a Python environment with pytest and access to the TileGym test suite.

## When your agent uses it

> Converts cuTile GPU kernels (@ct.kernel) to Triton (@triton.jit). Handles standard in-repo conversion, debugging (cudaErrorIllegalAddress, shape mismatch, numerical mismatch), and mapping cuTile idioms (ct.load/ct.store, ct.Constant, ct.launch) to Triton equivalents. Covers dual-kernel layout flags (e.g. transpose=True/False + autotune grid via META) per translations/advanced-patterns.md. Use when converting, porting, or translating cuTile kernels to Triton, or debugging existing Triton translations.

## What it can access

From a static check of the files on GitHub; nothing was run.

- [Runs Python scripts](https://github.com/NVIDIA/skills/tree/HEAD/skills/tilegym-converting-cutile-to-triton)
- [No network calls found in its scripts](https://github.com/NVIDIA/skills/tree/HEAD/skills/tilegym-converting-cutile-to-triton)
- [License: Apache-2.0](https://github.com/NVIDIA/skills/tree/HEAD/skills/tilegym-converting-cutile-to-triton)

## Features

- **Phased conversion workflow:** Steps through analysis, conversion, validation, testing and benchmarking with explicit gates between them.
- **cuTile to Triton API mapping:** Pairs cuTile idioms such as ct.load, ct.store, ct.Constant and ct.launch with their Triton equivalents.
- **Pre-flight source scan:** Grep commands that flag transposes, casts, launches and 2D loads in the cuTile source before any rewriting.
- **Runtime error debugging:** Guidance for illegal address errors, shape mismatches and numerical mismatches in translated kernels.
- **Performance regression checklist:** Lists patterns that commonly cause large slowdowns after conversion, with guidance on autotuning and benchmarking.
- **Layout flag handling:** Explains how to split dual-layout kernels into separate Triton kernels with a META-driven grid.

## Good for

- Port a cuTile attention kernel to Triton
- Debug an illegal address error in a Triton translation
- Convert a cuTile matmul with transpose flags into Triton kernels
- Check whether a converted Triton kernel is within 20% of cuTile performance
- Map ct.load and ct.store calls to Triton tensor descriptors

## Install

### Claude Code

Copy the skill folder to `~/.claude/skills/tilegym-converting-cutile-to-triton` to use it in all your projects:

```sh
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.claude/skills &&
rm -rf ~/.claude/skills/tilegym-converting-cutile-to-triton &&
cp -R "$tmp/skills/tilegym-converting-cutile-to-triton" ~/.claude/skills/tilegym-converting-cutile-to-triton &&
rm -rf "$tmp"
```

For one project only, put the folder in `.claude/skills/tilegym-converting-cutile-to-triton` inside it instead.

### Codex

Copy the skill folder to `~/.agents/skills/tilegym-converting-cutile-to-triton` to use it in every repository:

```sh
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.agents/skills &&
rm -rf ~/.agents/skills/tilegym-converting-cutile-to-triton &&
cp -R "$tmp/skills/tilegym-converting-cutile-to-triton" ~/.agents/skills/tilegym-converting-cutile-to-triton &&
rm -rf "$tmp"
```

For one repository only, put the folder in `.agents/skills/tilegym-converting-cutile-to-triton` inside it instead.

### Cursor

Copy the skill folder to `~/.cursor/skills/tilegym-converting-cutile-to-triton` to use it in every project:

```sh
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.cursor/skills &&
rm -rf ~/.cursor/skills/tilegym-converting-cutile-to-triton &&
cp -R "$tmp/skills/tilegym-converting-cutile-to-triton" ~/.cursor/skills/tilegym-converting-cutile-to-triton &&
rm -rf "$tmp"
```

For one project only, put the folder in `.cursor/skills/tilegym-converting-cutile-to-triton` inside it instead.

### Gemini CLI

Run in a terminal (add `--scope workspace` to install it for the current workspace only):

```sh
gemini skills install https://github.com/NVIDIA/skills.git --path skills/tilegym-converting-cutile-to-triton
```

Or copy the folder to `~/.gemini/skills/tilegym-converting-cutile-to-triton` yourself.

### VS Code

For GitHub Copilot in VS Code, copy the skill folder to `~/.copilot/skills/tilegym-converting-cutile-to-triton`:

```sh
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.copilot/skills &&
rm -rf ~/.copilot/skills/tilegym-converting-cutile-to-triton &&
cp -R "$tmp/skills/tilegym-converting-cutile-to-triton" ~/.copilot/skills/tilegym-converting-cutile-to-triton &&
rm -rf "$tmp"
```

For one project only, put the folder in `.github/skills/tilegym-converting-cutile-to-triton` inside it instead.

### Other agents

Get the folder:

```sh
git clone --depth 1 https://github.com/NVIDIA/skills.git
```

Then copy `skills/skills/tilegym-converting-cutile-to-triton` into your agent's skills folder as `tilegym-converting-cutile-to-triton`. Agents that support SKILL.md read one folder per skill.

## FAQ

### What does this skill change in my code?

It guides the agent through converting @ct.kernel functions to @triton.jit kernels and debugging existing Triton translations, following a checklist of conversion and validation steps.

### Which clients does it work with?

It is a plain SKILL.md folder that follows the shared agent skills format, so it applies to all supported clients.

### What does it expect as input?

It expects a cuTile kernel source file and the TileGym test suite, with pytest commands used for validation and benchmark scripts for performance comparison.

### Is it open source?

Yes, the skill is released under Apache-2.0 in the NVIDIA skills repository, with CC-BY-4.0 also listed in its metadata.

### How do I install Tilegym converting cutile to triton in Claude Code?

Copy the skill folder from GitHub to ~/.claude/skills/tilegym-converting-cutile-to-triton to use it in all your projects, or to .claude/skills/tilegym-converting-cutile-to-triton in one project.

### Does Tilegym converting cutile to triton run scripts?

Yes. It runs Python scripts. No network calls were found in its scripts. This comes from a static check of the files on GitHub; nothing was run. Read the files before you install it.

### Which AI agents can use Tilegym converting cutile to triton?

Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf and Goose. Each reads skills in the SKILL.md format from its own folder; the install steps on this page list them.

### What are the best alternatives to Tilegym converting cutile to triton?

The closest alternatives to Tilegym converting cutile to triton are Skill creator, Reverse engineer anything, Rea changelog update, Rea tool design and Rea analysis optimization.

## Where this comes from

- Description: from the skill's own SKILL.md.

## Alternatives

- [Skill creator](https://ai-powered.com/skills/skill-creator): Skill creator is a free agent skill for ai models and llm tooling.
- [Reverse engineer anything](https://ai-powered.com/skills/reverse-engineer-anything): Reverse engineer anything is an open-source agent skill for security.
- [Rea changelog update](https://ai-powered.com/skills/rea-changelog-update): Rea changelog update is an open-source agent skill for developer tools.
- [Rea tool design](https://ai-powered.com/skills/rea-tool-design): Rea tool design is an open-source agent skill for design and media.
- [Rea analysis optimization](https://ai-powered.com/skills/rea-analysis-optimization): Rea analysis optimization is an open-source agent skill for developer tools.
- [Rea ci tests audit](https://ai-powered.com/skills/rea-ci-tests-audit): Rea ci tests audit is an open-source agent skill for developer tools.
- [Diagnosing bugs](https://ai-powered.com/skills/diagnosing-bugs): Diagnosing bugs is an open-source agent skill for developer tools.
- [Domain modeling](https://ai-powered.com/skills/domain-modeling): Domain modeling is an open-source agent skill for developer tools.

---
Source: https://ai-powered.com/skills/tilegym-converting-cutile-to-triton
