Columbia University and CertiK researchers introduce TensorLean for GPU kernel verificationMachine translation
Columbia University 与 CertiK 研究人员推出 GPU 内核验证框架 TensorLean
TensorLean uses tensor-level math specifications written in Lean to verify compiled GPU kernels, checking for data races, termination, output within tolerance, and changes to other memory. The team checked 149 kernel instances, proved 83 correct, and found 38 correctness bugs. One FlashAttention-4 data race had not appeared in 17,760 test runs.
This source permits summary display only.
Read at the original source