SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference
This work improves efficient LLM inference for resource-constrained environments by compensating for KV cache compression, though it is incremental as it builds on existing eviction methods.
The paper tackled the problem of KV cache compression in LLMs by addressing saliency shift and marginal information over-compression, resulting in SmallKV achieving 1.75-2.56 times higher throughput than baselines while maintaining performance on benchmarks like GSM8K and LongBench.
KV cache eviction has emerged as an effective solution to alleviate resource constraints faced by LLMs in long-context scenarios. However, existing token-level eviction methods often overlook two critical aspects: (1) their irreversible eviction strategy fails to adapt to dynamic attention patterns during decoding (the saliency shift problem), and (2) they treat both marginally important tokens and truly unimportant tokens equally, despite the collective significance of marginal tokens to model performance (the marginal information over-compression problem). To address these issues, we design two compensation mechanisms based on the high similarity of attention matrices between LLMs of different scales. We propose SmallKV, a small model assisted compensation method for KV cache compression. SmallKV can maintain attention matching between different-scale LLMs to: 1) assist the larger model in perceiving globally important information of attention; and 2) use the smaller model's attention scores to approximate those of marginal tokens in the larger model. Extensive experiments on benchmarks including GSM8K, BBH, MT-Bench, and LongBench demonstrate the effectiveness of SmallKV. Moreover, efficiency evaluations show that SmallKV achieves 1.75 - 2.56 times higher throughput than baseline methods, highlighting its potential for efficient and performant LLM inference in resource constrained environments.