Researchers from Columbia University and CertiK have developed TensorLean, a formal verification framework designed to ensure the correctness of GPU kernels at the PTX level. This Lean-based system analyzes compiled kernels against tensor-level mathematical specifications, aiming to detect issues such as data races, termination failures, and incorrect outputs. In testing, TensorLean successfully verified 83 out of 149 kernel instances, uncovering 38 bugs across several projects, including significant issues within the FlashAttention suite. AI
IMPACT Enhances reliability and correctness of AI model training infrastructure by verifying GPU kernels.
RANK_REASON Academic paper detailing a new formal verification framework for GPU kernels. [lever_c_demoted from research: ic=1 ai=0.7]
- Certik
- Columbia University
- cvc5
- DeepGEMM
- FlashAttention-2
- FlashAttention-3
- FlashAttention-4
- GPU
- Lean
- Nvidia B200
- NVIDIA H100
- PyTorch
- TensorLean
- Z3
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →