KVCompose: Efficient Structured KV Cache Compression with Composite Tokens

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Akulov, Dmitry, Sana, Mohamed, De Domenico, Antonio, Salem, Tareq Si, Piovesan, Nicola, Ayed, Fadhel
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918144415629312
author Akulov, Dmitry
Sana, Mohamed
De Domenico, Antonio
Salem, Tareq Si
Piovesan, Nicola
Ayed, Fadhel
author_facet Akulov, Dmitry
Sana, Mohamed
De Domenico, Antonio
Salem, Tareq Si
Piovesan, Nicola
Ayed, Fadhel
contents Large language models (LLMs) rely on key-value (KV) caches for efficient autoregressive decoding; however, cache size grows linearly with context length and model depth, becoming a major bottleneck in long-context inference. Prior KV cache compression methods either enforce rigid heuristics, disrupt tensor layouts with per-attention-head variability, or require specialized compute kernels. We propose a simple, yet effective, KV cache compression framework based on attention-guided, layer-adaptive composite tokens. Our method aggregates attention scores to estimate token importance, selects head-specific tokens independently, and aligns them into composite tokens that respect the uniform cache structure required by existing inference engines. A global allocation mechanism further adapts retention budgets across layers, assigning more capacity to layers with informative tokens. This approach achieves significant memory reduction while preserving accuracy, consistently outperforming prior structured and semi-structured methods. Crucially, our approach remains fully compatible with standard inference pipelines, offering a practical and scalable solution for efficient long-context LLM deployment.
format Preprint
id arxiv_https___arxiv_org_abs_2509_05165
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle KVCompose: Efficient Structured KV Cache Compression with Composite Tokens
Akulov, Dmitry
Sana, Mohamed
De Domenico, Antonio
Salem, Tareq Si
Piovesan, Nicola
Ayed, Fadhel
Machine Learning
Large language models (LLMs) rely on key-value (KV) caches for efficient autoregressive decoding; however, cache size grows linearly with context length and model depth, becoming a major bottleneck in long-context inference. Prior KV cache compression methods either enforce rigid heuristics, disrupt tensor layouts with per-attention-head variability, or require specialized compute kernels. We propose a simple, yet effective, KV cache compression framework based on attention-guided, layer-adaptive composite tokens. Our method aggregates attention scores to estimate token importance, selects head-specific tokens independently, and aligns them into composite tokens that respect the uniform cache structure required by existing inference engines. A global allocation mechanism further adapts retention budgets across layers, assigning more capacity to layers with informative tokens. This approach achieves significant memory reduction while preserving accuracy, consistently outperforming prior structured and semi-structured methods. Crucially, our approach remains fully compatible with standard inference pipelines, offering a practical and scalable solution for efficient long-context LLM deployment.
title KVCompose: Efficient Structured KV Cache Compression with Composite Tokens
topic Machine Learning
url https://arxiv.org/abs/2509.05165