OBCache: Optimal Brain KV Cache Pruning for Efficient Long-Context LLM Inference

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Gu, Yuzhe, Liang, Xiyu, Zhao, Jiaojiao, Diao, Enmao
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913170881249280
author Gu, Yuzhe
Liang, Xiyu
Zhao, Jiaojiao
Diao, Enmao
author_facet Gu, Yuzhe
Liang, Xiyu
Zhao, Jiaojiao
Diao, Enmao
contents Large language models (LLMs) with extended context windows enable powerful 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. Code is available at https://github.com/DreamSoul-AI/OBCache.
format Preprint
id arxiv_https___arxiv_org_abs_2510_07651
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle OBCache: Optimal Brain KV Cache Pruning for Efficient Long-Context LLM Inference
Gu, Yuzhe
Liang, Xiyu
Zhao, Jiaojiao
Diao, Enmao
Computation and Language
Artificial Intelligence
Large language models (LLMs) with extended context windows enable powerful 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. Code is available at https://github.com/DreamSoul-AI/OBCache.
title OBCache: Optimal Brain KV Cache Pruning for Efficient Long-Context LLM Inference
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2510.07651