Anchor-based Precision Testing of Deep Learning Libraries

Deep learning operators are the fundamental computational units of model execution, and their numerical behavior critically depends on deep learning library implementations. Due to floating-point limitations and performance-oriented optimizations, operator implementations are prone to numerical precision issues, which may accumulate across layers and lead to degraded accuracy or inconsistent behaviors across libraries. While prior work has studied precision testing for deep learning libraries, most techniques compare only final model outputs, providing limited observability into where numerical discrepancies are introduced and how they evolve during execution.However, moving precision analysis from outputs to intermediate states is non-trivial. Numerical discrepancies are often introduced within long operator chains and amplified, attenuated, or masked by later computations. Moreover, intermediate tensors are high-dimensional and may be representationally misaligned across independent library implementations, making element-wise comparison brittle and prone to false positives. To fill this gap, this paper proposes an anchor-based precision testing approach that enables systematic detection and localization of numerical precision errors inside model execution. The approach instruments models with intermediate observation anchors at execution-stage boundaries to capture stage-level numerical states, and performs differential analysis using a consensus-based population deviation oracle to detect distributional deviations without relying on element-wise matching. To mitigate spurious discrepancies caused by implementation-induced representational differences, we further perform anchor-level alignment to reconcile semantically equivalent intermediate states across libraries.We evaluate ANPRED on YOLOv5 and YOLOv8 across three deep learning libraries against threshold-based output-level baselines. Results show that ANPRED not only preserves detection capability, but also localizes where precision issues emerge and how they propagate. Moreover, our experiments reveal that library-induced precision errors are stage-concentrated, dimension-sensitive, and non-monotonic in propagation. 

The relevant code has been uploaded in its entirety and is available in this repository. Due to space and size limitation, some typical data is uploaded, but complete data is supported, and the author can be contacted to obtain it if needed.