Large language models (LLMs) enable long-context tasks but face efficiency challenges due to the growing key-value (KV) cache. We propose LeanK, a learning-based method that prunes unimportant key (K) cache channels by leveraging static channel sparsity. With a novel two-stage training process, LeanK learns channel-wise static mask that could satisfy specific sparsity ratio and hardware alignment requirement. LeanK reduces GPU memory and accelerates decoding without sacrificing accuracy. Experiments demonstrate up to 70% K cache and 16%-18% V cache memory reduction. Custom decoding kernel enables 1.3x speedup for attention computation. We also provide insights into model channels and attention heads during long-context inference by analyzing the learned importance distribution. Our code is available at https://aka.ms/LeanK.
@article{arxiv.2508.02215,
title = {LeanK: Learnable K Cache Channel Pruning for Efficient Decoding},
author = {Yike Zhang and Zhiyuan He and Huiqiang Jiang and Chengruidong Zhang and Yuqing Yang and Jianyong Wang and Lili Qiu},
journal= {arXiv preprint arXiv:2508.02215},
year = {2025}
}