This release arrives following joint research from Columbia University and CertiK on November 22, 2026, targeting high-performance hardware architectures.
Developers can deploy TensorLean today to formally verify compiled production GPU kernels against mathematical specifications using Lean.
This release arrives following joint research from Columbia University and CertiK on November 22, 2026, targeting high-performance hardware architectures.
Verifying GPU Kernels at the PTX Level
A team of researchers from Columbia University and CertiK have introduced TensorLean, a framework designed for the formal verification of GPU kernels at the PTX level. TensorLean functions as a Lean-based verifier that mathematically proves compiled kernels match a designated tensor-level math specification.
The framework features a 1.8K-line PTX parser. Its specifications total 10,278 lines against 151,285 lines of kernel source, bringing the ratio to near 15:1. Operationally, the software runs on a 16-core workstation and checks PTX compiled for NVIDIA A100, H100, and B200 kernels.
During initial testing, the research team evaluated 149 kernel instances. This scope included every kernel of FlashAttention-2, FlashAttention-3, and FlashAttention-4, alongside DeepGEMM. Out of the 149 instances, the framework proved 83 correct. The remaining 66 instances led to 38 correctness bugs across six projects. Notably, the verification process caught a data race in FlashAttention-4 that had never appeared across 17,760 traditional test runs.
Confirmed Specifications and Open Factors
The parameters of the initial rollout and testing methodology are documented through formal specifications.
| Metric / Feature | Confirmed Detail |
|---|---|
| Framework Creators | Columbia University and CertiK |
| Release Date | November 22, 2026 |
| Target Architecture Level | PTX level (A100, H100, B200 kernels) |
| Tested Instances | 149 kernel instances (including FlashAttention-2, -3, -4, and DeepGEMM) |
| Verification Results | 83 proved correct, 38 bugs found across 6 projects |
| Hardware Requirement | 16-core workstation |
| Price or Version | Not in the sources |
| Unconfirmed Details | Not in the sources |
Price points and specific software version numbers remain unconfirmed in the current reporting.
Next Steps for Developers
Developers working with high-performance computing pipelines can examine the framework to audit production code for hidden concurrency issues and math discrepancies. Because TensorLean can isolate faults like the elusive FlashAttention-4 data race that standard test suites miss, engineering teams should evaluate the tool on local 16-core workstation setups running PTX compiled for supported NVIDIA architectures. Watch for subsequent updates from the research team regarding broader project coverage or expanded hardware compatibility.
Frequently Asked Questions
Q1: What is TensorLean?
TensorLean is a formal verification framework for GPU kernels at the PTX level created by researchers from Columbia University and CertiK.
Q2: What hardware and kernels does the framework check?
It runs on a 16-core workstation and checks PTX compiled for A100, H100, and B200 kernels, including every kernel of FlashAttention-2, -3, and -4 plus DeepGEMM.
Q3: How many bugs did the initial testing uncover?
The testing checked 149 kernel instances, proved 83 correct, and found 38 bugs across six projects.
