Finalist
Cappucino
Created a LLM-based-search for autotuner.
Core trick
An LLM scores candidate GPU kernel configurations with kernel source as context, and beam search keeps the top candidates.
Tags
Description
Created a LLM-based-search for autotuner. It takes in context of the kernel in scoring and performs a beam-search with the candidate kernel configurations.
Beats LFBO with a 1.43x speedup.
README excerpt
<div align="center"> <img src="docs/_static/helion_nobackground.png" alt="Helion Logo" width="250"/> </div> # About 📚 **[View Documentation](https://helionlang.com)** 📚 | 🎥 **[Watch Talk](https://youtu.be/BW-Ht-5IxgM)** 🎥 | 🚀 **[Try In Colab](https://colab.research.google.com/github/pytorch/helion/blob/main/notebooks/softmax.ipynb)** 🚀 | **[Try In AMD DevCloud](https://amd-ai-academy.com/github/pytorch/helion/blob/main/notebooks/softmax.ipynb)** **Helion** is a Python-embedded domain-specific language (DSL) for authoring machine learning kernels, designed to compile down to [Triton], a performant backend for programming GPUs and other devices. Helion aims to raise the level of abstraction compared to Triton, making it easier to write correct and efficient kernels while enabling more automation in the autotuning process. [Triton]: https://github.com/triton-lang/triton The name *Helion* refers to the nucleus of a helium-3 atom, while *Triton* refers to hydrogen-3. Helion can be viewed either as *PyTorch with tiles* or as *a higher-level Triton*. Compared to Triton, Helion reduces manual coding effort through autotuning. Helion spends more time (approx 10 min) autotuning as it evaluates hundreds of potential Triton implementations generated from a single Helion kernel. This larger search space also makes kernels more performance portable between different hardware. Helion automates and autotunes over: 1. **Tensor Indexing:** * Automatically calculates strides and in