LoopVQ: Sub-1-bit KV Cache Quantization for Looped Transformers
November 3, 2026
LoopVQ addresses the linear memory growth in looped transformers by applying vector quantization to the highly redundant KV cache. It leverages the varying sensitivity of different loop steps to optimize memory efficiency.
HOW THIS AFFECTS YOU
●
builderThis significantly reduces the memory bottleneck in looped transformer architectures, allowing for longer contexts.
●
researcherThis provides a method for optimizing parameter-efficient architectures through axis-specific quantization.