Toward Robust and Efficient ML-Based GPU Caching for Modern Inference

In modern GPU inference, cache efficiency remains a major bottleneck, and heuristic policies such as \textsc{LRU} can perform far worse than the offline optimum. Existing learning-based caching systems improve hit rates mainly through predictor design, but often follow learned predictions blindly, making performance unreliable when predictions are inaccurate. In contrast, emerging learning-augmented caching algorithms~\cite{pmlr-v80-lykouris18a,mitzenmacher2022algorithms} provide performance guarantees by carefully integrating predictions into caching policies, achieving both \emph{consistency} (near-optimality under perfect predictions) and \emph{robustness} (bounded worst-case performance under prediction errors). However, deployment remains challenging. A practical algorithm should satisfy strict time and space efficiency constraints, which some theoretical work overlooks, while also incurring low deployment overhead. We propose learning-augmented LRU, a deployment-oriented learning-augmented caching algorithm that guarantees \emph{1-consistency} and \emph{$O(k)$-robustness}, incurs low time and space overhead, and maintains strong compatibility. We further build a GPU cache, called \textsc{LCR}, on top of learning-augmented LRU to benefit from its theoretical guarantees and translate them into practical performance. In experiments, \textsc{LCR} reduces P99 time-to-first-token (TTFT) by up to 28.3\% on LLM workloads and increases throughput by up to 24.2\% on deep learning recommendation (DLRM) workloads. Even with poor predictions, performance degrades gracefully and remains close to \textsc{LRU}, demonstrating robustness with practical value.