OCTOPUS: Optimized KV Cache for Transformers via Octahedral Parametrization Under Optimal Squared Error Quantization

Researchers introduced OCTOPUS, a data-oblivious codec for compressing transformer KV caches.

It rotates coordinate triplets, uses octahedral parametrization, and applies optimized quantization to reduce error.

The method outperforms prior rotation-based codecs and can reconstruct keys on the fly without extra decode-time bandwidth or latency.

This matters because KV cache storage and bandwidth are major bottlenecks in long-context transformer inference.