November 22, 2026 · MarkTechPost
TensorLean formally verifies GPU kernels, finds 38 bugs
Researchers from Columbia University and CertiK built TensorLean, which uses the Lean proof assistant to verify production GPU kernels at the PTX (Nvidia's low-level instruction set) level. It checked 149 kernel instances, proved 83 correct, and found 38 bugs in libraries including FlashAttention and DeepGEMM.
Why it matters: Kernels like FlashAttention sit underneath most training and inference stacks, so silent bugs there can quietly corrupt results at scale. Verifying the compiled PTX rather than source code catches compiler-level issues too. Together with the Riemann proof check, it signals formal methods moving from academic niche into practical AI infrastructure.