Pyramid Cache: Layer-Adaptive KV Cache Compression with Signature-Based Cold Storage

Fuente: Zenodo
Enregistré dans:
Détails bibliographiques
Auteur principal: Sergio dj
Format: Recurso digital
Publié: Zenodo 2026
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866901841238818816
author Sergio dj
author_facet Sergio dj
contents <p class="p1">We present Pyramid Cache, a layer-adaptive compression architecture for transformer KV</p> <p class="p1">caches combined with a signature-based cold storage mechanism. Our approach exploits</p> <p class="p1">three properties: (1) token-identity structure varies systematically across layers, enabling</p> <p class="p1">layer-adaptive compression; (2) direction-only signatures predict attention relevance with</p> <p class="p1">0.926 Spearman correlation; and (3) only 0.4% of tokens receive meaningful attention at</p> <p class="p1">any step. A proof-of-concept implementation demonstrates that the system correctly</p> <p class="p1">answers both factual retrieval and complex reasoning questions — including multi-hop</p> <p class="p1">reasoning, causal chains, and contradiction detection — while skipping 99% of tokens</p> <p class="p1">(attending to only 5-8 out of 500+). Complex reasoning tasks show a similarity gap of only</p> <p class="p1">0.0215 compared to simple retrieval, with an EXCELLENT verdict at 1% reconstruction</p> <p class="p1">budget. Our approach is complementary to existing quantization methods like TurboQuant</p> <p class="p1">and represents a fundamentally different strategy: rather than compressing all vectors</p> <p class="p1">equally, we identify and skip the irrelevant ones entirely.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19536239
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Pyramid Cache: Layer-Adaptive KV Cache Compression with Signature-Based Cold Storage
Sergio dj
<p class="p1">We present Pyramid Cache, a layer-adaptive compression architecture for transformer KV</p> <p class="p1">caches combined with a signature-based cold storage mechanism. Our approach exploits</p> <p class="p1">three properties: (1) token-identity structure varies systematically across layers, enabling</p> <p class="p1">layer-adaptive compression; (2) direction-only signatures predict attention relevance with</p> <p class="p1">0.926 Spearman correlation; and (3) only 0.4% of tokens receive meaningful attention at</p> <p class="p1">any step. A proof-of-concept implementation demonstrates that the system correctly</p> <p class="p1">answers both factual retrieval and complex reasoning questions — including multi-hop</p> <p class="p1">reasoning, causal chains, and contradiction detection — while skipping 99% of tokens</p> <p class="p1">(attending to only 5-8 out of 500+). Complex reasoning tasks show a similarity gap of only</p> <p class="p1">0.0215 compared to simple retrieval, with an EXCELLENT verdict at 1% reconstruction</p> <p class="p1">budget. Our approach is complementary to existing quantization methods like TurboQuant</p> <p class="p1">and represents a fundamentally different strategy: rather than compressing all vectors</p> <p class="p1">equally, we identify and skip the irrelevant ones entirely.</p>
title Pyramid Cache: Layer-Adaptive KV Cache Compression with Signature-Based Cold Storage
url https://doi.org/10.5281/zenodo.19536239