TriAttention: Optimizing the KV Cache in LLMs TriAttention, a new KV cache compression technique, improves the efficiency of large language models by reducing memory use and increasing processing speed.