A Method for Building Large Language Models with Predefined KV Cache Capacity

This paper proposes a method for building large language models with predefined Key-Value (KV) cache capacity, particularly suitable for the attention layers in Transformer decode-only architectures. This method introduces fixed-length KV caches to address the issue of excessive memory consumption in traditional KV caches when handling infinite contexts. By dynamically updating the key-value vector sequences, it achieves efficient inference within limited cache capacity, significantly reducing memory usage while maintaining model performance and system throughput. Experimental results show that this method significantly reduces memory usage while maintaining the model's inference quality.