Finch: Prompt-guided Key-Value Cache Compression
This addresses memory and architectural constraints for applications like Retrieval-Augmented Generation and chatbots, though it is an incremental improvement on existing KV cache techniques.
The paper tackles the problem of processing long input contexts in large language models by proposing Finch, a method that compresses the Key-Value cache based on prompt relevance, achieving up to 93x compression while preserving semantic integrity without fine-tuning.
Recent large language model applications, such as Retrieval-Augmented Generation and chatbots, have led to an increased need to process longer input contexts. However, this requirement is hampered by inherent limitations. Architecturally, models are constrained by a context window defined during training. Additionally, processing extensive texts requires substantial GPU memory. We propose a novel approach, Finch, to compress the input context by leveraging the pre-trained model weights of the self-attention. Given a prompt and a long text, Finch iteratively identifies the most relevant Key (K) and Value (V) pairs over chunks of the text conditioned on the prompt. Only such pairs are stored in the KV cache, which, within the space constrained by the context window, ultimately contains a compressed version of the long text. Our proposal enables models to consume large inputs even with high compression (up to 93x) while preserving semantic integrity without the need for fine-tuning.