Codec-Gauge Optimizes Transformer KV Cache Compression.

As Transformer models handle increasingly long contexts, efficient compression and quantization of their Key-Value (KV) caches become critical. Existing research suggests that the channel basis of KV vectors significantly impacts how well compression backends preserve model behavior. This paper presents Codec-Gauge, a novel post-training layer designed to optimize this. Codec-Gauge learns small orthogonal channel transforms that reorient KV cache coordinates to be more "compression-friendly." Its objective function combines a token-channel DCT spectral-centroid loss with a smooth rate proxy, effectively concentrating KV energy into low-frequency layouts that are easier for codecs to handle. Evaluations across multiple models and bitrates demonstrate that learned gauges reduce zfp KL divergence by an average of 44% compared to raw coordinates and outperform other control methods. This approach also enhances quality preservation for block-uniform and KIVI-style quantization, establishing cache-coordinate geometry as a practical variable for improving compression fidelity without altering model weights or attention semantics.