Operator-Aware Mixed-Precision Tolerance Calibration for Tensor Kernels

A new calibration method for mixed-precision tolerance in tensor kernels, tested on the 26-entry gpuemu corpus with 8,076 result rows, achieves up to 2,184× tighter absolute tolerances than hand-picked values. For seven LLM-style buggy variants, the calibrated tolerances raise bug-detection recall from 73.2% to 82.4% (an absolute gain of 9.3 percentage points, +229 new detections) while increasing false positives from 0 to 20 out of 1,882 correct-control cases (+1.1 percentage points). The study, published on arXiv, mines element-wise error distributions from cloud GPU runs to determine the tolerance each kernel justifies under its correct implementation.

arXiv:2607.16228v1 Announce Type: new Abstract: Most tensor-kernel correctness tests go through a fixed-shape all close-style check with hand-picked absolute and relative tolerances. The thresholds are copied across the corpus and rarely revisited. We mine the element-wise error distribution of every test case from accumulated cloud GPU runs across the 26-entry gpuemu corpus and 2 dtypes 8,076 result rows . We then ask one empirical question: what absolute tolerance would the kernel itself, observed under its correct implementation, justify? The answer is much tighter than the current hand-picked atol. The largest tightening is attention triton fp16 at $2{,}184\times$. Restricted to the seven LLM-style buggy variants for which the corpus ships a paired correct counterpart, calibrated per- op, dtype tolerances raise bug-detection recall from 73.2% 1,805 of 2,467 to 82.4% 2,034 of 2,467 , an absolute gain of 9.3 percentage points +229 new detections . The control false-positive count rises from 0 to 20 out of 1,882 correct-control cases +1.1 percentage points .