[Submitted connected 13 Aug 2026]
View PDF HTML (experimental)
Abstract:Systems that make GPU kernels pinch connection models study precocious correctness rates. Those rates travel from a azygous loose test: tally the kernel connected a fewer random inputs astatine 1 fixed style and judge it if the output is adjacent to a reference. A kernel tin walk that trial and still beryllium silently wrong. It tin return an mean number wherever the existent reply is simply a NaN aliases an infinity, disagree from tally to run, break erstwhile the style changes, aliases accumulate successful fp16 wherever the reference keeps an fp32 total. We build the instrumentality that checks correctness properly: a contract-grade verifier of 12 adversarial gates, each a spot a correct kernel must satisfy, respective of them tolerance-free, truthful nary prime of period tin explicate a nonaccomplishment away. Aimed outward, the verifier audits 2,638 machine-generated kernels that a nationalist system's ain harness had already accepted arsenic correct. It finds 39.5% surgery beyond immoderate tolerance statement and 62.1% carrying astatine slightest 1 violation. The field's modular trial accepts 1,487 kernels the verifier rejects, against only 14 the different way. We take sides the uncovering 4 independent ways: a 7/7 affirmative control, a threshold-calibration sweep, 98.5% statement pinch the reference benchmark's ain correctness code, and a stratified hand-audit. Aimed inward, the verifier judges a kernel of our own: the first autochthonal Blackwell tcgen05 training backward for the gated-linear-recurrence (GDN) family, including the reverse-state shape the section still runs connected a fallback. We found its correctness independently, against a double-precision oracle, and train 5 family members done it. The correctness awesome down reported advancement successful kernel procreation is acold weaker than the numbers suggest, and a group of tolerance-free contracts would adjacent astir of the gap.Submission history
From: Rishi Shah [view email]
[v1] Thu, 13 Aug 2026 01:25:56 UTC (149 KB)
English (US) ·
Indonesian (ID) ·