-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
248 lines (216 loc) · 11.3 KB
/
index.html
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
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<!-- Meta tags for social media banners, these should be filled in appropriatly as they are your "business card" -->
<!-- Replace the content tag with appropriate information -->
<meta name="description" content="xKV: Cross-Layer SVD for KV-Cache Compression">
<meta property="og:title" content="xKV"/>
<meta property="og:description" content="xKV: Cross-Layer SVD for KV-Cache Compression"/>
<meta property="og:url" content="https://abdelfattah-lab.github.io/xKV/"/>
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
<meta property="og:image" content="https://abdelfattah-lab.github.io/xKV/static/images/overview_small.jpg" />
<meta property="og:image:width" content="1200"/>
<meta property="og:image:height" content="561"/>
<meta name="twitter:title" content="xKV">
<meta name="twitter:description" content="xKV: Cross-Layer SVD for KV-Cache Compression">
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600-->
<meta name="twitter:image" content="https://abdelfattah-lab.github.io/xKV/static/images/overview_small.jpg">
<meta name="twitter:card" content="summary_large_image">
<!-- Keywords for your paper to be indexed by-->
<meta name="keywords" content="KV Cache">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>xKV: Cross-Layer SVD for KV-Cache Compression-</title>
<link rel="icon" type="image/x-icon" href="static/images/xKV_logo_cute.ico">
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro"
rel="stylesheet">
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="static/css/bulma-slider.min.css">
<link rel="stylesheet" href="static/css/fontawesome.all.min.css">
<link rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="static/css/index.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<script src="static/js/bulma-carousel.min.js"></script>
<script src="static/js/bulma-slider.min.js"></script>
<script src="static/js/index.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="is-size-5 publication-authors">
<h1 class="title is-2 publication-title"><img src="static/images/xKV_logo_cute.png" alt="xKV" width="50" height="50" style="display: inline; vertical-align: bottom;" /> xKV: Cross-Layer SVD for KV-Cache Compression</h1>
<div class="is-size-5 publication-authors">
<!-- Paper authors -->
<span class="author-block">
<a href="FIRST AUTHOR PERSONAL LINK" target="_blank">Chi-Chih Chang</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="SECOND AUTHOR PERSONAL LINK" target="_blank">Chien-Yu Lin</a><sup>2</sup>,
</span>
<span class="author-block">
<a href="THIRD AUTHOR PERSONAL LINK" target="_blank">Yash Akhauri</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="THIRD AUTHOR PERSONAL LINK" target="_blank">Wei-Cheng Lin</a><sup>3</sup>,
</span><br>
<span class="author-block">
<a href="THIRD AUTHOR PERSONAL LINK" target="_blank">Kai-Chiang Wu</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="THIRD AUTHOR PERSONAL LINK" target="_blank">Luis Ceze</a><sup>2</sup>,
</span>
<span class="author-block">
<a href="THIRD AUTHOR PERSONAL LINK" target="_blank">Mohamed S. Abdelfattah </a><sup>1</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="affliation"><small><sup>1</sup>Cornell University, <sup>2</sup>University of Washington, </span><br> <sup>3</sup>National Yang Ming Chiao Tung University</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- Github link -->
<span class="link-block">
<a href="https://github.com/abdelfattah-lab/xKV.git" target="_blank"
class="external-link button is-normal is-rounded is-dark"
style="color: #fff !important;"
>
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<!-- ArXiv abstract Link -->
<span class="link-block">
<a href="https://arxiv.org/abs/2503.18893" target="_blank"
class="external-link button is-normal is-rounded is-dark"
style="color: #fff !important;"
>
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full">
<!-- Title with icon -->
<h4 class="title is-4">
<img src="static/images/target.png" style="height: 40px; display: inline; vertical-align: middle;"/>
Abstract
</h4>
<!-- The figure block -->
<div class="figure" style="text-align: center;">
<img src="static/images/teaser.png"
alt=""
style="max-width: 100%; height: auto;" />
</div>
<div class="level-set has-text-justified">
<p>
Large Language Models (LLMs) with long context windows enable powerful applications but come at the cost of high memory consumption to store the Key and Value states (KV-Cache). Recent efforts to compress KV-Cache—such as token eviction, quantization, and low-rank decomposition—have primarily focused on intra-layer redundancies that compress each layer’s KV-Cache independently. While this often yields respectable per-layer compression, these methods do not exploit potential redundancy across layers. Moreover, attempts to merge the KV-cache from multiple layers typically require expensive training or rely on assumptions of high per-token cosine similarity, which does not generally hold in practice.
</p><br>
<p>
In this work, we identify that the dominant singular vectors are remarkably well-aligned across multiple layers of the KV-Cache. Leveraging this alignment, we introduce <strong>xKV</strong>, a simple post-training method that applies SVD on the KV-Cache of a group of layers. Our method consolidates the KV-Cache of multiple layers into a shared low-rank subspace, yielding a more compact KV-Cache representations.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section hero is-light is-small" style="padding: 0rem 1.5rem;">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full">
<!-- Title with icon -->
<h4 class="title is-4">
<img src="static/images/startup.png" style="height: 43px; display: inline; vertical-align: middle;"/>
Highlights
</h4>
<!-- Bullet Points -->
<div class="content" style="margin-bottom: 1rem;">
<ul style="list-style-type: disc; margin-left: 1.5em;">
<li><strong>Up to 8× KV-Cache reduction</strong>
on widely-used LLMs (e.g., Llama-3.1 and Qwen2.5)
while maintaining accuracy.
</li>
<li><strong>3× compression rate</strong> on coding tasks
for Multi-Head Latent Attention (MLA) models
(e.g., DeepSeek-Coder-V2) without performance degradation.
</li>
</ul>
</div>
<div class="figure" style="text-align: center;">
<img src="static/images/ruler.jpg"
alt="Accuracy comparison"
style="max-width: 90%; height: auto; border-radius: 8px;" />
<p class="subtitle has-text-centered" style="font-size: 1rem; margin-top: 0.75em;">
Accuracy comparison of MiniCache, applying SVD on single layer’s KV-Cache
and xKV (cross-layer SVD) on Llama-3.1-8B-Instruct (left)
and Qwen2.5-14B-Instruct-1M (right). Results are averaged across tasks from RULER benchmark.
</p>
</div>
<br>
<div class="figure" style="text-align: center;">
<img src="static/images/deepseek-coder.jpg"
alt="DeepSeek evaluation"
style="max-width: 90%; height: auto; border-radius: 8px;" />
<p class="subtitle has-text-centered" style="font-size: 1rem; margin-top: 0.75em;">
Evaluation results of different KV-Cache methods
on DeepSeek-Coder-V2-Lite-Instruct model.
</p>
<br>
</div>
</div>
</div>
</div>
</section>
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{chang2024palu,
title={xKV: Cross-Layer SVD for KV-Cache Compression},
author={Chang, Chi-Chih and Lin, Chien-Yu and Akhauri, Yash and Lin, Wei-Cheng and Wu, Kai-Chiang and Ceze, Luis and Abdelfattah, Mohamed S.},
year={2025},
journal={arXiv preprint arXiv:2503.18893},
year={2025}
}</code></pre>
</div>
</section>
<!--End BibTex citation -->
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a> which was adopted from the <a href="https://nerfies.github.io" target="_blank">Nerfies</a> project page.
You are free to borrow the source code of this website, we just ask that you link back to this page in the footer. This website is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">Creative
Commons Attribution-ShareAlike 4.0 International License</a>. Icons created by <a href="https://www.flaticon.com">Smashicons - Flaticon</a>
</p>
</div>
</div>
</div>
</div>
</footer>
<!-- Statcounter tracking code -->
<!-- You can add a tracker to track page visits by creating an account at statcounter.com -->
<!-- End of Statcounter Code -->
</body>
</html>