Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: O'Neill, James, Clancy, Robert, Matskevichus, Mariia, Reid, Fergal
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2601.11471
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866908944822173696
author O'Neill, James
Clancy, Robert
Matskevichus, Mariia
Reid, Fergal
author_facet O'Neill, James
Clancy, Robert
Matskevichus, Mariia
Reid, Fergal
contents The key-value (KV) cache is a primary memory bottleneck in Transformers. We propose Low-Rank Key-Value (LRKV) attention, which reduces KV cache memory by exploiting redundancy across attention heads, while being compute efficient. Each layer uses a shared full-rank KV projection augmented with low-rank, head-specific residuals, providing a continuous trade-off between complete sharing and full independence. After pretraining models of size 128M to 6.3B parameters, LRKV consistently achieves the lowest test loss among standard MHA, MQA/GQA, and MLA while using only 45-53\% of MHA's KV cache. LRKV reaches equivalent baseline quality 18-25\% faster (measured in training steps). After supervised midtraining, LRKV achieves the highest downstream task performance across ARC-Easy, ARC-Challenge, MMLU, GSM8K, and HumanEval benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2601_11471
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Low-Rank Key Value Attention
O'Neill, James
Clancy, Robert
Matskevichus, Mariia
Reid, Fergal
Machine Learning
The key-value (KV) cache is a primary memory bottleneck in Transformers. We propose Low-Rank Key-Value (LRKV) attention, which reduces KV cache memory by exploiting redundancy across attention heads, while being compute efficient. Each layer uses a shared full-rank KV projection augmented with low-rank, head-specific residuals, providing a continuous trade-off between complete sharing and full independence. After pretraining models of size 128M to 6.3B parameters, LRKV consistently achieves the lowest test loss among standard MHA, MQA/GQA, and MLA while using only 45-53\% of MHA's KV cache. LRKV reaches equivalent baseline quality 18-25\% faster (measured in training steps). After supervised midtraining, LRKV achieves the highest downstream task performance across ARC-Easy, ARC-Challenge, MMLU, GSM8K, and HumanEval benchmarks.
title Low-Rank Key Value Attention
topic Machine Learning
url https://arxiv.org/abs/2601.11471