Catching NaN at the MLIR Pass Boundary (thecloudlet.github.io)

A recent investigation into the Multi-Level Intermediate Representation (MLIR) reveals that NaN (not-a-number) values can emerge unexpectedly during the constant folding process in compiler passes, even when all operands are valid. This can lead to significant accuracy regressions in machine learning models. The discovery serves as a wake-up call for the AI/ML community, emphasizing the need for rigorous internal checks within compilers. By implementing a pass-boundary verification system, the compiler can catch these erroneous NaN values before they propagate to the runtime, ensuring higher reliability in computational results. The technical implications are substantial; the problem often arises when operations such as multiplying zero by infinity yield NaN, despite producing structurally valid intermediate representations. The solution proposed includes enhancing the MLIR framework with local operation verifiers that check for specific numeric invariants associated with operations. By integrating these verification checks, the compiler can maintain the integrity of numeric attributes throughout the transformation process, preventing faulty values from compromising model accuracy. The initiative highlights a crucial step towards building more robust AI systems, where accuracy and reliability in computations are paramount.