Wink Pings

Columbia University and CertiK Launch TensorLean, 149 GPU Kernels Tested Reveal 38 Hidden Bugs

Research teams from Columbia University and security firm CertiK have launched TensorLean, a tool based on the Lean language, to conduct formal verification on production-grade GPU kernels at the PTX instruction layer. The test covered 149 deployed kernel instances: a full logical correctness proof was completed for 83 instances, and 38 real bugs were found in mainstream AI computing components such as FlashAttention and DeepGEMM, providing a new verification approach for the reliability of underlying computing power for large models.

Columbia University and CertiK's research team have jointly launched TensorLean, a formal verification tool for GPU kernels. The related results were published on November 21, 2026.

First, two basic concepts are added for readers without technical backgrounds.

GPU kernels are low-level computing codes running on graphics cards. Core operations such as attention calculation and matrix multiplication during large model training and inference are directly executed by corresponding kernels. In the past, the industry's troubleshooting for kernel problems mainly relied on sampling test cases, which was difficult to cover all edge operating scenarios. Some hidden bugs may lurk for months before being triggered, causing minor deviations in calculation results or even directly interrupting training tasks that take weeks to complete.

PTX is NVIDIA's general intermediate instruction set, located between upper-level application code and lower-level hardware instructions. TensorLean chooses to perform verification at this layer, so no repeated adaptation is required for different models of graphics cards. The tool is developed based on the theorem-proving language Lean, adopting a formal verification approach that confirms the full-scenario correctness of code logic through mathematical deductions.

This time, the team used TensorLean to test 149 GPU kernel instances in production. Finally, full correctness proofs were completed for 83 instances, and 38 existing code bugs were identified, involving widely used AI computing components such as FlashAttention and DeepGEMM. Many open-source large model frameworks have built-in kernel codes for the above components.

Formal verification was previously more applied in scenarios with extremely low fault tolerance rates, such as aerospace embedded codes and blockchain smart contracts. As large models gradually land in key fields such as finance and medical care, the reliability requirements of underlying computing will continue to improve, and the landing scenarios of such verification tools will continue to expand.

Source: MarkTechPost, Author: Michal Sutter

发布时间: 2026-11-22 09:51