{"id":"https://openalex.org/W4404088015","doi":"https://doi.org/10.48550/arxiv.2410.14731","title":"MatryoshkaKV: Adaptive KV Compression via Trainable Orthogonal Projection","display_name":"MatryoshkaKV: Adaptive KV Compression via Trainable Orthogonal Projection","publication_year":2024,"publication_date":"2024-10-16","ids":{"openalex":"https://openalex.org/W4404088015","doi":"https://doi.org/10.48550/arxiv.2410.14731"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2410.14731","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.14731","pdf_url":"https://arxiv.org/pdf/2410.14731","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2410.14731","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5047724144","display_name":"Bokai Lin","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Lin, Bokai","raw_affiliation_strings":["Shanghai Jiao Tong University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5043552523","display_name":"Zihao Zeng","orcid":"https://orcid.org/0009-0009-9243-7368","observed_orcids":["https://orcid.org/0009-0009-9243-7368","https://orcid.org/0000-0002-4745-5331"]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Zeng, Zihao","raw_affiliation_strings":["Shanghai Jiao Tong University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Xiao, Zipeng","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Xiao, Zipeng","raw_affiliation_strings":["Shanghai Jiao Tong University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5024119845","display_name":"Siqi Kou","orcid":"https://orcid.org/0009-0008-4073-9818","observed_orcids":["https://orcid.org/0009-0008-4073-9818"]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Kou, Siqi","raw_affiliation_strings":["Shanghai Jiao Tong University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5060320373","display_name":"TianQi Hou","orcid":"https://orcid.org/0000-0003-1940-4833","observed_orcids":["https://orcid.org/0000-0003-1940-4833"]},"institutions":[{"id":"https://openalex.org/I2250955327","display_name":"Huawei Technologies (China)","ror":"https://ror.org/00cmhce21","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250955327"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Hou, Tianqi","raw_affiliation_strings":["Huawei"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Huawei","institution_ids":["https://openalex.org/I2250955327"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5019439900","display_name":"Xiaofeng Gao","orcid":"https://orcid.org/0000-0003-1776-8799","observed_orcids":["https://orcid.org/0000-0003-1776-8799","https://orcid.org/0000-0002-8361-1573"]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Gao, Xiaofeng","raw_affiliation_strings":["Shanghai Jiao Tong University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Zhang, Hao","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I36258959","display_name":"University of California San Diego","ror":"https://ror.org/0168r3w48","country_code":"US","type":"education","lineage":["https://openalex.org/I36258959"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zhang, Hao","raw_affiliation_strings":["University of California , San Diego {19821172068 , zengzihao , xiaozp 25 ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of California , San Diego {19821172068 , zengzihao , xiaozp 25 ,","institution_ids":["https://openalex.org/I36258959"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5102623510","display_name":"Zhijie Deng","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Deng, Zhijie","raw_affiliation_strings":["Shanghai Jiao Tong University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University","institution_ids":["https://openalex.org/I183067930"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":19,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.6309000253677368,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.6309000253677368,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.27469998598098755,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.0284000001847744,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/kv-cache-compression","display_name":"KV cache compression","score":0.9819999933242798},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8180000185966492},{"id":"https://openalex.org/keywords/orthogonal-projection","display_name":"orthogonal projection","score":0.781000018119812},{"id":"https://openalex.org/keywords/low-rank-projection","display_name":"low-rank projection","score":0.8759999871253967},{"id":"https://openalex.org/keywords/dimensionality-reduction","display_name":"dimensionality reduction","score":0.5690000057220459},{"id":"https://openalex.org/keywords/principal-component-analysis","display_name":"principal component analysis","score":0.7049999833106995},{"id":"https://openalex.org/keywords/knowledge-distillation","display_name":"knowledge distillation","score":0.9739999771118164},{"id":"https://openalex.org/keywords/self-attention","display_name":"self-attention","score":0.7059999704360962}],"concepts":[{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.6516196727752686},{"id":"https://openalex.org/C57493831","wikidata":"https://www.wikidata.org/wiki/Q3134666","display_name":"Projection (relational algebra)","level":2,"score":0.5831252336502075},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.47337931394577026},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.25792405009269714},{"id":"https://openalex.org/C192562407","wikidata":"https://www.wikidata.org/wiki/Q228736","display_name":"Materials science","level":0,"score":0.2133232057094574},{"id":"https://openalex.org/C159985019","wikidata":"https://www.wikidata.org/wiki/Q181790","display_name":"Composite material","level":1,"score":0.06916427612304688}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2410.14731","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.14731","pdf_url":"https://arxiv.org/pdf/2410.14731","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2410.14731","doi":"https://doi.org/10.48550/arxiv.2410.14731","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2410.14731","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2410.14731","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.14731","pdf_url":"https://arxiv.org/pdf/2410.14731","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G5394961855","display_name":"\u57fa\u4e8e\u8d1d\u53f6\u65af\u6df1\u5ea6\u5b66\u4e60\u7684\u504f\u5fae\u5206\u65b9\u7a0b\u89e3\u7b97\u5b50\u5b66\u4e60","funder_award_id":"23ZR1428700","funder_id":"https://openalex.org/F4320309612","funder_display_name":"Natural Science Foundation of Shanghai"}],"funders":[{"id":"https://openalex.org/F4320309612","display_name":"Natural Science Foundation of Shanghai","ror":null},{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4404088015.pdf","grobid_xml":"https://content.openalex.org/works/W4404088015.grobid-xml"},"referenced_works_count":26,"referenced_works":["https://openalex.org/W1821462560","https://openalex.org/W2794325560","https://openalex.org/W2950700477","https://openalex.org/W2981852735","https://openalex.org/W2988394319","https://openalex.org/W3030163527","https://openalex.org/W3168867926","https://openalex.org/W4319653544","https://openalex.org/W4386185625","https://openalex.org/W4387294617","https://openalex.org/W4387561528","https://openalex.org/W4389518760","https://openalex.org/W4389650671","https://openalex.org/W4391591383","https://openalex.org/W4392090070","https://openalex.org/W4392537434","https://openalex.org/W4395065284","https://openalex.org/W4399414358","https://openalex.org/W4399598242","https://openalex.org/W4403749471","https://openalex.org/W4411403346","https://openalex.org/W4415796681","https://openalex.org/W4415798413","https://openalex.org/W4415798938","https://openalex.org/W4415799217","https://openalex.org/W7133224090"],"related_works":["https://openalex.org/W4403885701","https://openalex.org/W7164021007","https://openalex.org/W7134049577","https://openalex.org/W7140325981","https://openalex.org/W4415436488","https://openalex.org/W4401306371","https://openalex.org/W4402671912","https://openalex.org/W4404088887","https://openalex.org/W7125409367","https://openalex.org/W4417410215"],"abstract_inverted_index":{"KV":[0,46,226],"cache":[1,31,47,76,98,227],"has":[2],"become":[3,50],"a":[4,51,153,194],"de":[5],"facto":[6],"technique":[7],"for":[8,34,78,114,167,172,240],"the":[9,37,40,45,54,70,75,86,97,109,124,141,148,168,205],"inference":[10],"of":[11,18,39,74,209,230],"large":[12],"language":[13],"models":[14],"(LLMs),":[15],"where":[16,129],"tensors":[17,77],"shape":[19],"(layer":[20],"number,":[21,23],"head":[22],"sequence":[24],"length,":[25],"feature":[26,87],"dimension)":[27],"are":[28],"introduced":[29],"to":[30,95,145,182,234],"historical":[32],"information":[33],"self-attention.":[35],"As":[36],"size":[38],"model":[41],"and":[42,59,175,192,199,213,245],"data":[43,115,207],"grows,":[44],"can":[48,187,218],"quickly":[49],"bottleneck":[52],"within":[53],"system":[55],"in":[56,236],"both":[57],"storage":[58],"memory":[60],"transfer.":[61],"To":[62,139],"address":[63],"this,":[64],"prior":[65],"studies":[66],"usually":[67],"focus":[68],"on":[69,85],"first":[71],"three":[72],"axes":[73],"compression.":[79],"This":[80],"paper":[81],"supplements":[82],"them,":[83],"focusing":[84],"dimension":[88],"axis,":[89],"by":[90,107],"utilizing":[91],"low-rank":[92],"projection":[93,112,128,150],"matrices":[94,151],"transform":[96],"features":[99],"into":[100],"spaces":[101],"with":[102,126,152,223],"reduced":[103],"dimensions.":[104],"We":[105,122,202],"begin":[106],"investigating":[108],"canonical":[110],"orthogonal":[111,149],"method":[113,186,217],"compression":[116,137,170,179,200,228],"through":[117],"principal":[118],"component":[119],"analysis":[120],"(PCA).":[121],"observe":[123],"issue":[125],"PCA":[127],"significant":[130],"performance":[131,198,222],"degradation":[132],"is":[133],"observed":[134],"at":[135],"low":[136],"rates.":[138],"bridge":[140],"gap,":[142],"we":[143,164],"propose":[144],"directly":[146],"tune":[147],"distillation":[154],"objective":[155],"using":[156],"an":[157,224],"elaborate":[158],"Matryoshka":[159],"training":[160,211],"strategy.":[161],"After":[162],"training,":[163],"adaptively":[165],"search":[166],"optimal":[169],"rates":[171],"various":[173],"layers":[174],"heads":[176],"given":[177],"varying":[178],"budgets.":[180],"Compared":[181],"previous":[183],"works,":[184],"our":[185,210,216],"easily":[188],"embrace":[189],"pre-trained":[190],"LLMs":[191,242],"hold":[193],"smooth":[195],"tradeoff":[196],"between":[197],"rate.":[201],"empirically":[203],"witness":[204],"high":[206],"efficiency":[208],"procedure":[212],"find":[214],"that":[215],"sustain":[219],"over":[220],"90%":[221],"average":[225],"rate":[229],"60%":[231],"(and":[232],"up":[233],"75%":[235],"certain":[237],"extreme":[238],"scenarios)":[239],"popular":[241],"like":[243],"LLaMA2-7B-base":[244],"Mistral-7B-v0.3-base.":[246]},"counts_by_year":[{"year":2026,"cited_by_count":11},{"year":2025,"cited_by_count":5},{"year":2024,"cited_by_count":3}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2024-11-06T00:00:00"}