elastickv/CITATION.cff at main · infolake/elastickv · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
cff-version: 1.2.0
message: If you use ElasticKV, please cite this software release using the metadata below.
title: "ElasticKV: Exploiting Attention Haze for Structure-Preserving KV-Cache Compression"
type: software
version: 0.1.1
date-released: 2026-04-11
doi: 10.5281/zenodo.19503351
license: Apache-2.0
authors:
  - family-names: Camargo
    given-names: Guilherme de
    affiliation: Independent Researcher
    orcid: "https://orcid.org/0009-0004-8913-9419"
contact:
  - family-names: Camargo
    given-names: Guilherme de
    affiliation: Independent Researcher
    email: "camargo@phiq.io"
repository-code: "https://github.com/infolake/elastickv"
url: "https://github.com/infolake/elastickv"
abstract: >-
  ElasticKV is a training-free KV-cache sparsification method for llama.cpp
  that removes low-magnitude, low-selectivity attention haze via per-head
  adaptive thresholds with sink token protection. This archival release bundles
  the CUDA hook, paper source, raw artifacts, and analysis scripts used in the
  accompanying manuscript.
identifiers:
  - type: doi
    value: 10.5281/zenodo.19503351
    description: Concept DOI covering all Zenodo versions of ElasticKV.
keywords:
  - kv-cache
  - llama.cpp
  - llm
  - cuda
  - long-context
  - sparsity
  - attention
  - reproducibility