Reproduce “EGG: An Expert-Guided Agent Framework for Kernel Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yaochen Han, Ke Fan, Hongxu Jiang, Wanqi Xu, Weiyu Xie, Runhua Zhang, Chenhui Zhu, Yixiang Zhang
Researchers at Beihang University and collaborating institutions developed EGG, an expert-guided agent framework that generates high-performance GPU kernels by integrating expert optimization principles and a multi-agent collaboration mechanism. The system achieved a 100% success rate and an average 2.13x speedup over PyTorch Eager, even outperforming some hand-tuned production kernels.
Key insights
- Decomposing the complex kernel generation into expert-guided hierarchical stages allows LLMs to tackle the problem more effectively and efficiently.
- Integrating domain-specific runtime profiling and debugging feedback within a multi-agent loop is crucial for achieving high correctness and targeted performance improvements.
- Explicitly managing and propagating optimization context across stages prevents objective drift and enables a progressive, cumulative improvement in kernel performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.