English | 简体中文 | 繁體中文 | 日本語 | Français

TileGym

TileGym is a CUDA Tile kernel library that provides a rich collection of kernel tutorials and examples for tile-based GPU programming.

Overview | Features | Installation | Quick Start | Contributing | License

Overview

This repository aims to provide helpful kernel tutorials and examples for tile-based GPU programming. TileGym is a playground for experimenting with CUDA Tile, where you can learn how to build efficient GPU kernels and explore their integration into real-world large language models such as Llama 3.1 and DeepSeek V2. Whether you're learning tile-based GPU programming or looking to optimize your LLM implementations, TileGym offers practical examples and comprehensive guidance.

Features

• Rich collection of CUDA Tile kernel examples
• Practical kernel implementations for common deep learning operators
• Performance benchmarking to evaluate kernel efficiency
• End-to-end integration examples with popular LLMs (Llama 3.1, DeepSeek V2)

Installation

Prerequisites

GPU Support: TileGym requires CUDA 13.1+ and a Blackwell GPU (e.g., B200, RTX 5080, RTX 5090). NVIDIA Ampere (e.g., A100) is also supported with CUDA 13.2+. All released kernels are validated on both architectures. Download CUDA from NVIDIA CUDA Downloads.

• PyTorch (version 2.9.1 or compatible)
• CUDA 13.1+ (Required - TileGym is built and tested exclusively on CUDA 13.1+)
• Triton (included with PyTorch installation)

Setup Steps

1. Prepare torch and triton environment

If you already have torch and triton, skip this step.

pip install --pre torch --index-url https://download.pytorch.org/whl/cu130

We have verified that torch==2.9.1 works. You can also get triton packages when installing torch.

2. Install TileGym

TileGym uses cuda-tile (≥ 1.5.0) for GPU kernel programming, which depends on the tileiras compiler at runtime.

Install from PyPI (recommended)

pip install tilegym[tileiras]

This installs TileGym and all runtime dependencies, including cuda-tile[tileiras] which bundles the tileiras compiler directly into your Python environment.

If you already have tileiras available on your system (e.g., from CUDA Toolkit 13.1+), you can omit the extra:

pip install tilegym

Install from source

git clone https://github.com/NVIDIA/TileGym.git
cd TileGym
pip install .[tileiras]   # or: pip install .  (if you have system tileiras)

For editable (development) mode, use pip install -e . or pip install -e .[tileiras].

All runtime dependencies are declared in requirements.txt and are installed automatically by both pip install tilegym and pip install ..

We also provide Dockerfile, you can refer to modeling/transformers/README.md.

Backends

TileGym provides kernels for the following backends, each in its own folder under src/tilegym/ops/:

• cuTile (default) — src/tilegym/ops/cutile, see more details in cutile-python.
• CUDA Tile C++ — src/tilegym/ops/tilecpp, see more details in README.tilecpp.md.
• Triton CUDA Tile IR — src/tilegym/ops/triton, see more details in Triton-to-tile-IR.

To use the Triton CUDA Tile IR backend, install its wheel into a separate directory and select it at runtime with ENABLE_TILE=1. Wheels for CPython 3.12 and 3.13 are available on the releases page:

# Install into a separate directory, kept apart from the default environment
pip install --target /opt/triton_tileir <triton-tileir-wheel-for-your-python>.whl

# Select the Triton CUDA Tile IR backend at runtime
PYTHONPATH=/opt/triton_tileir ENABLE_TILE=1 python your_script.py

Quick Start

There are three main ways to use TileGym:

1. Explore Kernel Examples

All kernel implementations are located in the src/tilegym/ops/ directory. You can test individual operations with minimal scripts. Function-level usage and minimal scripts for individual ops are documented in tests/ops/README.md

2. Run Benchmarks

Evaluate kernel performance with micro-benchmarks:

cd tests/benchmark
bash run_all.sh

Complete benchmark guide available in tests/benchmark/README.md

3. Run LLM Transformer Examples

Use TileGym kernels in end-to-end inference scenarios. We provide runnable scripts and instructions for transformer language models (e.g., Llama 3.1-8B) accelerated using TileGym kernels.

First, install the additional dependency:

pip install accelerate==1.13.0 --no-deps

Containerized Setup (Docker):

docker build -t tilegym-transformers -f modeling/transformers/Dockerfile .
docker run --gpus all -it tilegym-transformers bash

More details in modeling/transformers/README.md

4. Julia (cuTile.jl) Kernels (Optional)

TileGym also includes experimental cuTile.jl kernel implementations in Julia. These are self-contained in the julia/ directory and do not require the Python TileGym package.

Prerequisites: Julia 1.12+, CUDA 13.1, Blackwell GPU

# Install Julia (if not already installed)
curl -fsSL https://install.julialang.org | sh

# Install dependencies
julia --project=julia/ -e 'using Pkg; Pkg.instantiate()'

# Run tests
julia --project=julia/ julia/test/runtests.jl

See julia/Project.toml for the full dependency list.

5. Enable the cuTile-rs (Rust) backend (Optional)

A subset of ops ships an additional cuTile-rs backend under src/tilegym/ops/cutile_rs — kernels authored in Rust with cutile-rs and loaded through a C-ABI libcutile_kernels.so. It is opt-in and only usable from a source checkout.

Prerequisites (in addition to the base install above), matching cuTile-rs:

• Rust 1.89+ — cargo and rustc on PATH:

  curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
  rustup default stable

• CUDA toolkit with headers — the Rust build runs bindgen against cuda.h. Set CUDA_TOOLKIT_PATH to your install; if unset, cuTile-rs falls back to /usr/local/cuda:

  export CUDA_TOOLKIT_PATH=/usr/local/cuda   # must contain include/cuda.h

Use it. The backend loader builds the shared library lazily on first use (cargo build --release), so no manual build step is required:

import tilegym
tilegym.set_backend("cutile-rs")

from tilegym.backend.selector import get_available_backends
print(get_available_backends())        # should include "cutile-rs"

from tilegym.ops import bmm             # backend-agnostic import
# ... bmm(...) now dispatches to the cuTile-rs kernel

Optional environment knobs:

export CUTILE_RS_AUTOBUILD=0                          # skip the lazy rebuild; use a prebuilt .so
export CUTILE_RS_KERNELS_DIR=/abs/path/to/cutile_kernels   # override the crate location

If cargo is not on PATH and no prebuilt libcutile_kernels.so is present, the backend reports itself unavailable and cuTile-rs tests are skipped rather than failing.

Benchmarking cuTile-rs. When comparing cuTile-rs perf against the cuTile-Python baseline, run the perf tests with CUPTI=1 (uses CUPTI / torch.profiler device time instead of CUDA events). cuTile-rs kernels often have different host/launch overhead than the reference, which CUDA-event wall timing over-counts on small (sub-microsecond) kernels; CUPTI measures pure GPU kernel time and gives a stable, apples-to-apples ratio:

CUPTI=1 pytest tests/ops/test_bmm.py -k "test_perf and cutile_rs" --print-record

Contributing

We welcome contributions of all kinds. Please read our CONTRIBUTING.md for guidelines, including the Contributor License Agreement (CLA) process.

License and third-party notices

• Project license: MIT
  • LICENSE
• Third-party attributions and license texts:
  • LICENSES/ATTRIBUTIONS.md