Formally Verifying AI-Generated GPU Kernels (gimletlabs.ai)

Gimlet Labs has introduced a pioneering approach for formal verification of AI-generated GPU kernels, addressing a critical bottleneck in the production trustworthiness of these optimized codes. While AI agents have rapidly improved in generating efficient GPU kernels, confidence in their outputs remains a challenge due to potential bugs that might not surface in traditional numerical tests. Current testing methods often overlook subtle discrepancies that can lead to significant functional issues in real workloads, which can manifest under specific conditions that aren't captured during standard testing. To tackle this, Gimlet Labs is developing a tensor algebra equivalence checker that uses formal verification to ensure semantic equivalence between reference PyTorch models and their AI-optimized implementations, specifically those employing Triton kernels. By translating both reference and candidate kernels into mathematical formulas, the system uses SMT solvers to verify correctness across all possible inputs, thus significantly enhancing confidence in kernel outputs. This method not only aims to catch bugs that numeric testing misses but also aligns with Jason Wei's "Verifier's Law,” emphasizing the necessity of robust verification techniques as AI continues to evolve in application areas like GPU optimization.