Reproduce “AKG kernel Agent: A Multi-Agent Framework for Cross-Platform Kernel Synthesis” 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
Jinye Du, Quan Yuan, Zuyao Zhang, Yanzhi Yi, Jiahui Hu, Wangyi Chen, Yiyang Zhu, Qishui Zheng and 11 more
The AKG system uses four agents to generate processor-specific kernels from operator specifications: a Designer creates a portable optimization sketch, a Coder translates it into target-specific code, and a Verifier and Conductor test and route failures for revision. On KernelBench Level 1, Triton-Ascend reports a 1.46× geometric mean speedup over PyTorch Eager on evaluated kernels; this measures execution time, and profiling remains pending.
Key insights
- Decoupling optimization strategy and high-level design through a hardware-agnostic Unified Sketch simplifies the task for code generation LLMs and enhances portability across diverse targets.
- A multi-agent, closed-loop framework with dynamic error analysis and task routing (Conductor) enables adaptive refinement and targeted debugging for complex kernel synthesis.
- Document-Driven Integration (DDI) and hierarchical code retrieval effectively provide relevant, structured knowledge to LLMs, facilitating extensibility and performance across new DSLs and hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.