DynaCalKV: Key-Value Cache Compression via Head Grouping and Adaptive Rank Allocation

As the inference phase of Large Language Models (LLMs) requires handling long context windows, the Key-Value (KV) cache initially appears to address this challenge but eventually becomes a significant bottleneck as the context window continues to grow. Low-rank compression has recently been studied ...