
from amd-agi
Numerical correctness verification for Primus-Turbo operators.
This skill provides a rigorous framework for proving that a Primus-Turbo operator is numerically correct by comparing it against a higher-precision PyTorch reference for both forward and backward passes.
float32 (FP32) to validate lower-precision outputs.allclose tolerances (e.g., $1\text{e-}4$ for FP32).tests/pytorch/ref/.pytest -n 8.This skill has not been reviewed by our automated audit pipeline yet.