Скачать или смотреть triton_lite, a Triton clone in Mojo: Jeff Niu at the Modular GPU Kernel Hackathon

In this talk, Jeff Niu from OpenAI explores how he brought Triton-style abstractions into Mojo to prototype high-performance kernels ahead of the Modular GPU Kernel Hackathon. Drawing on his experience working on both the Triton compiler and the Mojo language, Jeff shares how he ported Triton's value-semantic tile abstractions into Mojo and built a Python integration for testing against PyTorch and Triton references.

He walks through the benefits and tradeoffs of this approach, highlights how the type system in Mojo enables layout propagation, and discusses how this design could support future tooling like torch.compile. While acknowledging performance limitations compared to Triton, Jeff presents a compelling case for Mojo as a flexible, expressive platform for kernel development and experimentation.

Recorded at the Modular GPU Kernel Hackathon at AGI House on May 10, 2025.

Catch the hackathon highlight reel:    • Mojo meets AMD MI300X: Modular GPU Kernel ...  
Check out Chris Lattner's opening remarks:    • Democratizing AI compute together: Chris L...  
Full hackathon playlist:    • Democratizing AI compute together: Chris L...  

Join our community 🤝:
Forum - https://forum.modular.com/
GitHub - https://github.com/modular
X (aka Twitter) - https://x.com/modular
LinkedIn -   / modular-ai  
Reddit -   / modularai  

#ModularHackathon #MojoLang #AMD #Crusoe #GPUMODE #AI