CertiK and Columbia University Bring Math Proofs to the GPU Code Behind Every Large AI Model
4 Articles
4 Articles
A Columbia and CertiK framework uses Lean and SMT tools to check GPU kernels at PTX level. In 149 instances analyzed, the team demonstrated that 83 were compliant with the specified properties and reported 38 errors in six projects, including a data career in FlashAttention-4 that, according to researchers, did not appear in 17.760 test executions. *** TensorLean evaluated 149 kernel instances and demonstrated that 83 were compliant with the spe…
A Team of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
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.
CertiK and Columbia University Unveil TensorLean, Identify 25 Bugs in AI GPU Kernels
Summary CertiK and Columbia University introduced TensorLean, a mathematical verification framework designed to detect computational errors in production AI GPU kernels. Researchers examined over 100 GPU kernels, uncovering 25 software bugs, including two in DeepSeek and 23 in FlashAttention implementations. TensorLean enables developers to verify optimized GPU calculations, reducing risks of costly training repetitions and errors affecting AI m…
Coverage Details
Bias Distribution
- 100% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium








