In-context KV-Cache Eviction for LLMs via Attention-Gate
This addresses efficiency issues in LLM inference systems, though it is an incremental improvement over existing KV-Cache techniques.
The paper tackles the KV-Cache bottleneck in LLM inference by introducing a dynamic eviction policy using Attention-Gates, which improves efficiency and performance by caching only a subset of tokens.
The KV-Cache technique has become the standard for the inference of large language models (LLMs). Yet, it is widely criticized that KV-Cache can become a bottleneck of the LLM inference system. This paper enables a novel dynamic KV-Cache eviction policy by injecting a lightweight module called Attention-Gate to the model. It accepts the global context as input and yields eviction flags for each token. The self-attention modules in the model proceed according to the flags and cache only a subset of the KV states for next token prediction. The Attention-Gates can yield various flags for different heads and layers and be easily tuned on top of a pre-trained LLM via continual pre-training or supervised fine-tuning. The computational and memory overhead introduced by Attention-Gates can be minimal. We empirically evaluate the proposed approach across multiple scenarios, showing that effective eviction of redundant tokens can not only improve efficiency but also enhance performance.