Reproduce “The Output-Space Hypothesis: Enumerative Equivalence Checking for Tensor Programs” 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
Paul Biberstein, Joseph Devietti, Mayur Naik
Dirigo checks whether optimized GPU programs match references by symbolically tracing mixed PyTorch/CUDA tensor programs, then asking an SMT solver whether a chosen output can differ for any inputs. Among 6,988 KernelBench level-2 programs passing differential testing, it confirmed 600 buggy programs, with 97.3% found within two minutes. It does not prove all programs correct and supports restricted models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.