Tilus is a powerful domain-specific language (DSL) for GPU programming that offers:
- Thread-block-level granularity with tensors as the primary data type.
- Explicit control over shared memory and register tensors (unlike Triton).
- Low-precision types with arbitrary bit-widths (1 to 8 bits).
It also includes automatic tuning, caching, and a Pythonic interface for ease of use.
Tilus is pronounced as tie-lus, /ˈtaɪləs/.
Install Tilus using pip:
pip install tilus
Note
Tilus depends on cuda-python. If your GPU driver is older than 580.65.06, you will need to install an older version of cuda-python to ensure compatibility.
pip install tilus "cuda-python<13"
To get started, refer to the tutorials to learn how to program kernels with Tilus.
You can also check more examples of using Tilus.
You can learn more on different topics in the programming guide.
This project is based on the following research paper:
@article{ding2025tilus,
title={Tilus: A Virtual Machine for Arbitrary Low-Precision GPGPU Computation in LLM Serving},
author={Ding, Yaoyao and Hou, Bohan and Zhang, Xiao and Lin, Allan and
Chen, Tianqi and Hao, Cody Yu and Wang, Yida and Pekhimenko, Gennady},
journal={arXiv preprint arXiv:2504.12984},
year={2025}
}We would like to acknowledge the following projects for their influence on Tilus's design and development:
- Hidet: We take Hidet IR as our low-level target and reuse its runtime system.
- TVM: Hidet's initial IR was adopted from TVM, and we also learned a lot from TVM on how to build a compiler.
- Triton: The core idea of defining kernels at a thread-block level and working with tiles was inspired by Triton.
- Hexcute: We adopted the idea of using automatic layout inference to simplify programming from Hexcute.
