D-Quant: Driftable Entropy Coding for KV Cache Quantization
September 18, 2026
D-Quant introduces driftable entropy coding to address memory bottlenecks in LLM deployment caused by KV cache growth. It moves beyond fixed-width quantization to better exploit the non-uniform, normal distribution of KV values after rotation and normalization.
HOW THIS AFFECTS YOU
●
builderYou can reduce the memory footprint and bandwidth pressure of your LLM deployments by using non-fixed-width quantization for the KV cache.