TensorLean, from Columbia University and CertiK, formally verifies production GPU kernels at the PTX level using Lean. It checked 149 kernel instances, proved 83 correct, and found 38 bugs in FlashAttention, DeepGEMM and more.
Read the original article:
