CRVQ: Channel-relaxed Vector Quantization for Extreme Compression of LLMs

Powerful large language models (LLMs) are increasingly expected to be deployed with lower computational costs, enabling their capabilities on resource-constrained devices. Post-training quantization (PTQ) has emerged as a star approach to achieve this ambition, with best methods compressing weights...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Xu, Yuzhuang, Ji, Shiyu, Zhu, Qingfu, Che, Wanxiang
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Xu, Yuzhuang
Ji, Shiyu
Zhu, Qingfu
Che, Wanxiang
description Powerful large language models (LLMs) are increasingly expected to be deployed with lower computational costs, enabling their capabilities on resource-constrained devices. Post-training quantization (PTQ) has emerged as a star approach to achieve this ambition, with best methods compressing weights to less than 2 bit on average. In this paper, we propose Channel-Relaxed Vector Quantization (CRVQ), a novel technique that significantly improves the performance of PTQ baselines at the cost of only minimal additional bits. This state-of-the-art extreme compression method achieves its results through two key innovations: (1) carefully selecting and reordering a very small subset of critical weight channels, and (2) leveraging multiple codebooks to relax the constraint of critical channels. With our method, we demonstrate a 38.9% improvement over the current strongest sub-2-bit PTQ baseline, enabling nearer lossless 1-bit compression. Furthermore, our approach offers flexible customization of quantization bit-width and performance, providing a wider range of deployment options for diverse hardware platforms.
doi_str_mv 10.48550/arxiv.2412.09282
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2412_09282</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2412_09282</sourcerecordid><originalsourceid>FETCH-arxiv_primary_2412_092823</originalsourceid><addsrcrecordid>eNpjYJA0NNAzsTA1NdBPLKrILNMzMjE00jOwNLIw4mTwdA4KC7RScM5IzMtLzdEtSs1JrEhNUQhLTS7JL1IILE3MK8msSizJzM9TSAMKuFaUFKXmpio45-cWFKUWF4PE89MUfHx8i3kYWNMSc4pTeaE0N4O8m2uIs4cu2NL4gqLM3MSiyniQ5fFgy40JqwAA3Xo5cg</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>CRVQ: Channel-relaxed Vector Quantization for Extreme Compression of LLMs</title><source>arXiv.org</source><creator>Xu, Yuzhuang ; Ji, Shiyu ; Zhu, Qingfu ; Che, Wanxiang</creator><creatorcontrib>Xu, Yuzhuang ; Ji, Shiyu ; Zhu, Qingfu ; Che, Wanxiang</creatorcontrib><description>Powerful large language models (LLMs) are increasingly expected to be deployed with lower computational costs, enabling their capabilities on resource-constrained devices. Post-training quantization (PTQ) has emerged as a star approach to achieve this ambition, with best methods compressing weights to less than 2 bit on average. In this paper, we propose Channel-Relaxed Vector Quantization (CRVQ), a novel technique that significantly improves the performance of PTQ baselines at the cost of only minimal additional bits. This state-of-the-art extreme compression method achieves its results through two key innovations: (1) carefully selecting and reordering a very small subset of critical weight channels, and (2) leveraging multiple codebooks to relax the constraint of critical channels. With our method, we demonstrate a 38.9% improvement over the current strongest sub-2-bit PTQ baseline, enabling nearer lossless 1-bit compression. Furthermore, our approach offers flexible customization of quantization bit-width and performance, providing a wider range of deployment options for diverse hardware platforms.</description><identifier>DOI: 10.48550/arxiv.2412.09282</identifier><language>eng</language><subject>Computer Science - Computation and Language ; Computer Science - Learning</subject><creationdate>2024-12</creationdate><rights>http://creativecommons.org/licenses/by/4.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,776,881</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2412.09282$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2412.09282$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Xu, Yuzhuang</creatorcontrib><creatorcontrib>Ji, Shiyu</creatorcontrib><creatorcontrib>Zhu, Qingfu</creatorcontrib><creatorcontrib>Che, Wanxiang</creatorcontrib><title>CRVQ: Channel-relaxed Vector Quantization for Extreme Compression of LLMs</title><description>Powerful large language models (LLMs) are increasingly expected to be deployed with lower computational costs, enabling their capabilities on resource-constrained devices. Post-training quantization (PTQ) has emerged as a star approach to achieve this ambition, with best methods compressing weights to less than 2 bit on average. In this paper, we propose Channel-Relaxed Vector Quantization (CRVQ), a novel technique that significantly improves the performance of PTQ baselines at the cost of only minimal additional bits. This state-of-the-art extreme compression method achieves its results through two key innovations: (1) carefully selecting and reordering a very small subset of critical weight channels, and (2) leveraging multiple codebooks to relax the constraint of critical channels. With our method, we demonstrate a 38.9% improvement over the current strongest sub-2-bit PTQ baseline, enabling nearer lossless 1-bit compression. Furthermore, our approach offers flexible customization of quantization bit-width and performance, providing a wider range of deployment options for diverse hardware platforms.</description><subject>Computer Science - Computation and Language</subject><subject>Computer Science - Learning</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNpjYJA0NNAzsTA1NdBPLKrILNMzMjE00jOwNLIw4mTwdA4KC7RScM5IzMtLzdEtSs1JrEhNUQhLTS7JL1IILE3MK8msSizJzM9TSAMKuFaUFKXmpio45-cWFKUWF4PE89MUfHx8i3kYWNMSc4pTeaE0N4O8m2uIs4cu2NL4gqLM3MSiyniQ5fFgy40JqwAA3Xo5cg</recordid><startdate>20241212</startdate><enddate>20241212</enddate><creator>Xu, Yuzhuang</creator><creator>Ji, Shiyu</creator><creator>Zhu, Qingfu</creator><creator>Che, Wanxiang</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20241212</creationdate><title>CRVQ: Channel-relaxed Vector Quantization for Extreme Compression of LLMs</title><author>Xu, Yuzhuang ; Ji, Shiyu ; Zhu, Qingfu ; Che, Wanxiang</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-arxiv_primary_2412_092823</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Computer Science - Computation and Language</topic><topic>Computer Science - Learning</topic><toplevel>online_resources</toplevel><creatorcontrib>Xu, Yuzhuang</creatorcontrib><creatorcontrib>Ji, Shiyu</creatorcontrib><creatorcontrib>Zhu, Qingfu</creatorcontrib><creatorcontrib>Che, Wanxiang</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Xu, Yuzhuang</au><au>Ji, Shiyu</au><au>Zhu, Qingfu</au><au>Che, Wanxiang</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>CRVQ: Channel-relaxed Vector Quantization for Extreme Compression of LLMs</atitle><date>2024-12-12</date><risdate>2024</risdate><abstract>Powerful large language models (LLMs) are increasingly expected to be deployed with lower computational costs, enabling their capabilities on resource-constrained devices. Post-training quantization (PTQ) has emerged as a star approach to achieve this ambition, with best methods compressing weights to less than 2 bit on average. In this paper, we propose Channel-Relaxed Vector Quantization (CRVQ), a novel technique that significantly improves the performance of PTQ baselines at the cost of only minimal additional bits. This state-of-the-art extreme compression method achieves its results through two key innovations: (1) carefully selecting and reordering a very small subset of critical weight channels, and (2) leveraging multiple codebooks to relax the constraint of critical channels. With our method, we demonstrate a 38.9% improvement over the current strongest sub-2-bit PTQ baseline, enabling nearer lossless 1-bit compression. Furthermore, our approach offers flexible customization of quantization bit-width and performance, providing a wider range of deployment options for diverse hardware platforms.</abstract><doi>10.48550/arxiv.2412.09282</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2412.09282
ispartof
issn
language eng
recordid cdi_arxiv_primary_2412_09282
source arXiv.org
subjects Computer Science - Computation and Language
Computer Science - Learning
title CRVQ: Channel-relaxed Vector Quantization for Extreme Compression of LLMs
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-25T22%3A28%3A15IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=CRVQ:%20Channel-relaxed%20Vector%20Quantization%20for%20Extreme%20Compression%20of%20LLMs&rft.au=Xu,%20Yuzhuang&rft.date=2024-12-12&rft_id=info:doi/10.48550/arxiv.2412.09282&rft_dat=%3Carxiv_GOX%3E2412_09282%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true