English

ParisKV: Fast and Drift-Robust KV-Cache Retrieval for Long-Context LLMs

Machine Learning 2026-02-11 v2 Computation and Language Databases

Abstract

KV-cache retrieval is essential for long-context LLM inference, yet existing methods struggle with distribution drift and high latency at scale. We introduce ParisKV, a drift-robust, GPU-native KV-cache retrieval framework based on collision-based candidate selection, followed by a quantized inner-product reranking estimator. For million-token contexts, ParisKV supports CPU-offloaded KV caches via Unified Virtual Addressing (UVA), enabling on-demand top-kk fetching with minimal overhead. ParisKV matches or outperforms full attention quality on long-input and long-generation benchmarks. It achieves state-of-the-art long-context decoding efficiency: it matches or exceeds full attention speed even at batch size 1 for long contexts, delivers up to 2.8×\times higher throughput within full attention's runnable range, and scales to million-token contexts where full attention runs out of memory. At million-token scale, ParisKV reduces decode latency by 17×\times and 44×\times compared to MagicPIG and PQCache, respectively, two state-of-the-art KV-cache Top-kk retrieval baselines.

Keywords

Cite

@article{arxiv.2602.07721,
  title  = {ParisKV: Fast and Drift-Robust KV-Cache Retrieval for Long-Context LLMs},
  author = {Yanlin Qi and Xinhang Chen and Huiqiang Jiang and Qitong Wang and Botao Peng and Themis Palpanas},
  journal= {arXiv preprint arXiv:2602.07721},
  year   = {2026}
}

Comments

25 pages, 16 figures. Under review

R2 v1 2026-07-01T10:26:18.497Z