• Sources: primary
  • Summary: The paper derives per-element acceptance regions from precision-annotated reference programs, tracking shared rounding, block-scaled quantizers and math library error. On 13 Triton operators across two GPU generations it raised no false alarms on 33,960 correct runs per GPU, while tuned tolerances flagged up to 30 percent of correct runs under input shift. It killed 116 of 124 mutants on an RTX PRO 6000 versus 59 to 83 for fixed baselines.
  • Why it matters: Kernel authors testing FP8 and block-scaled code get an alternative to hand-tuned tolerances that either flake or miss real accuracy bugs.