Kernel Contracts: A Specification Language for ML Kernel Correctness Across Heterogeneous Silicon
Researchers propose a formal specification language for ML kernel contracts that defines what kernels should compute, resolving ambiguities when the same operation produces different results across AMD, NVIDIA, and other platforms. The framework covers precision, ordering, compiler, and exceptional-value failures, with a three-state calibration requirement and demonstrated application to real incidents like Huawei's silent precision coercion and AMD's out-of-bounds handling.
arXiv ML · 3 min (abstract)
Research