XC-Cache: Cross-Attending to Cached Context for Efficient LLM Inference

TL;DR

This paper introduces XC-Cache, a method that uses cross-attention to efficiently condition large language models on reference text without extensive caching, reducing memory usage and improving performance over traditional in-context learning.

Contribution

The paper proposes a novel cross-attention based approach inspired by encoder-decoder models, enabling efficient conditioning of decoder-only models without prompt-based caching.

Findings

Outperforms traditional in-context learning in QA tasks

Reduces cache space by two orders of magnitude

Achieves performance comparable to fine-tuned models

Abstract

In-context learning (ICL) approaches typically leverage prompting to condition decoder-only language model generation on reference information. Just-in-time processing of a context is inefficient due to the quadratic cost of self-attention operations, and caching is desirable. However, caching transformer states can easily require almost as much space as the model parameters. When the right context isn't known in advance, caching ICL can be challenging. This work addresses these limitations by introducing models that, inspired by the encoder-decoder architecture, use cross-attention to condition generation on reference text without the prompt. More precisely, we leverage pre-trained decoder-only models and only train a small number of added layers. We use Question-Answering (QA) as a testbed to evaluate the ability of our models to perform conditional generation and observe that they outperform ICL, are comparable to fine-tuned prompted LLMs, and drastically reduce the space footprint relative to standard KV caching by two orders of magnitude.