Researchers from Columbia University and CertiK have introduced TensorLean, a framework for formal verification of GPU kernels at the PTX level. TensorLean is a Lean-based verifier that proves compiled kernels match a tensor-level math spec. The research team checked 149 kernel instances, including every kernel of FlashAttention-2, -3 and -4 plus DeepGEMM. It proved 83 correct. The remaining 66 led to 38 correctness bugs.
TL;DR
- Size: 1.8K-line PTX parser. Specs total 10,278 lines against 151,285 lines of kernel source, a ratio near 15:1.
- Runs on: A 16-core workstation. It checks PTX compiled for A100, H100 and B200 kernels.
- Performance: 83 of 149 kernel instances proved correct, 38 bugs found across 6 projects.
- Best: Caught a FlashAttention-4 data race that never appeared in 17,760 test runs.
What is TensorLean?
TensorLean is a formal verification framework for GPU kernels at the PTX level.
It starts from unmodified source code. A developer writes a tensor-level spec in Lean. TensorLean then proves 4 properties: no data races, termination, correct output within tolerance, and untouched memory elsewhere.
How does TensorLean verify a GPU kernel?
Specs build on Lean’s Mathlib. Tensor operators follow PyTorch names, so @ means torch.matmul. Each spec has 5 parts: spec_math, spec_mut, spec_layout, spec_tol and requires. The math stays separate from the memory layout.
The pipeline runs in 7 steps:
- Start from unmodified source.
- Write the tensor-level spec.
- Compile to PTX and load a PTX semantics.
- Build a happens-before graph to check races and termination.
- Symbolically execute to get final memory, identical across schedules.
- Lower the spec to PTX-memory level.
- Discharge proofs with Lean’s kernel, Z3 and cvc5.
Which techniques make it scale?
- Concurrency to sequential: Race-freedom proofs on the happens-before graph cover all CTAs, threads and loop iterations symbolically.
- Loop invariants over semirings: Closed forms for tiled, accumulating loops, such as online softmax, need 1 SMT query per loop.
- Tolerance proofs: Algebraic error is proved zero. Round-off error is bounded separately, then combined.
- Assertions as hints: Developers add
tensorlean_assertin source. Each is proved once, then reused.
What bugs did TensorLean find?
The GitHub README lists 38 bugs, all reported upstream. FlashAttention accounts for 23. Liger Kernel has 7, TileOPs 3, TileLang 2, DeepGEMM 2 and Triton 1.
By category, the site reports 19 memory access errors and 13 numerical bugs. It also counts 7 data races, 5 uninitialized memory reads and 5 undefined behaviors. There are 3 algorithm errors, 2 miscompilations and 1 deadlock. One bug can sit in several categories.
Of 102 FlashAttention instances, 56 passed all 4 properties. The other 46 each failed a proof goal on a source bug.


