Reproduce “LARK: Lean-Assisted Reasoning for LLM-Generated GPU Kernels” 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
Zixi Zhang, Jiayi Nie, Haoran Wu, Yujie Yan, Yann Herklotz, Rika Antonova, Robert Mullins, Yiren Zhao and 2 more
LARK uses Lean to certify low-level GPU kernels against formalized PyTorch specifications. It checks race freedom, verifies that kernel segmentation preserves the flat model, and proves segment-wise equivalence under exact arithmetic. The proofs rely on trusted task interfaces and machine semantics, and do not establish host-allocation bounds. LARK certified nine of ten adapted production kernels, rejecting one for a race.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.