A Team of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
A Team of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
Researchers from Columbia University and CertiK introduced TensorLean, a Lean-based framework that formally verifies GPU kernels at the PTX level against tensor-level math specs, proving no data races, termination, correct output within tolerance, and untouched memory elsewhere. Across 149 kernel instances, including all kernels of FlashAttention-2, -3 and -4 plus DeepGEMM, it proved 83 correct and found 38 bugs across 6 projects, including a FlashAttention-4 data race that never appeared in 17,760 test runs. Specs total 10,278 lines against 151,285 lines of kernel source, and it runs on a 16-core workstation checking PTX compiled for A100, H100 and B200 kernels.
Source: MarkTechPost · marktechpost.com