OBCache: Optimal Brain KV Cache Pruning for Efficient Long-Context LLM Inference
This addresses memory efficiency for LLM inference, but it is incremental as it enhances existing cache eviction methods with output-aware signals.
The paper tackles the memory overhead in large language models with long contexts by proposing OBCache, a framework that formulates cache eviction as structured pruning using Optimal Brain Damage theory, resulting in improved long-context accuracy in experiments on LLaMA and Qwen models.
Large language models (LLMs) with extended context windows enable powerful downstream applications but impose significant memory overhead, as caching all key-value (KV) states scales linearly with sequence length and batch size. Existing cache eviction methods address this by exploiting attention sparsity, yet they typically rank tokens heuristically using accumulated attention weights without considering their true impact on attention outputs. We propose Optimal Brain Cache (OBCache), a principled framework that formulates cache eviction as a layer-wise structured pruning problem. Building upon the Optimal Brain Damage (OBD) theory, OBCache quantifies token saliency by measuring the perturbation in attention outputs induced by pruning tokens, with closed-form scores derived for isolated keys, isolated values, and joint key-value pairs. Our scores account not only for attention weights but also for information from value states and attention outputs, thereby enhancing existing eviction strategies with output-aware signals. Experiments on LLaMA and Qwen models demonstrate that replacing the heuristic scores in existing works, which estimate token saliency across different query positions, with OBCache's output-aware scores consistently improves long-context accuracy.