{"id":"https://openalex.org/W4402727767","doi":"https://doi.org/10.1109/mwscas60917.2024.10658840","title":"Hessian-Aware KV Cache Quantization for LLMs","display_name":"Hessian-Aware KV Cache Quantization for LLMs","publication_year":2024,"publication_date":"2024-08-11","ids":{"openalex":"https://openalex.org/W4402727767","doi":"https://doi.org/10.1109/mwscas60917.2024.10658840"},"language":"en","primary_location":{"id":"doi:10.1109/mwscas60917.2024.10658840","is_oa":false,"landing_page_url":"https://doi.org/10.1109/mwscas60917.2024.10658840","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2024 IEEE 67th International Midwest Symposium on Circuits and Systems (MWSCAS)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":false,"oa_status":"closed","oa_url":null,"any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5109786813","display_name":"Woohong Byun","orcid":null},"institutions":[{"id":"https://openalex.org/I130701444","display_name":"Georgia Institute of Technology","ror":"https://ror.org/01zkghx44","country_code":"US","type":"education","lineage":["https://openalex.org/I130701444"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Woohong Byun","raw_affiliation_strings":["School of Electrical and Computer Engineering, Georgia Institute of Technology,Atlanta,Georgia,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Electrical and Computer Engineering, Georgia Institute of Technology,Atlanta,Georgia,USA","institution_ids":["https://openalex.org/I130701444"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5080060539","display_name":"Jongseok Woo","orcid":"https://orcid.org/0000-0001-8030-2356"},"institutions":[{"id":"https://openalex.org/I130701444","display_name":"Georgia Institute of Technology","ror":"https://ror.org/01zkghx44","country_code":"US","type":"education","lineage":["https://openalex.org/I130701444"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Jongseok Woo","raw_affiliation_strings":["School of Electrical and Computer Engineering, Georgia Institute of Technology,Atlanta,Georgia,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Electrical and Computer Engineering, Georgia Institute of Technology,Atlanta,Georgia,USA","institution_ids":["https://openalex.org/I130701444"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5009591041","display_name":"Saibal Mukhopadhyay","orcid":"https://orcid.org/0000-0002-8894-3390"},"institutions":[{"id":"https://openalex.org/I130701444","display_name":"Georgia Institute of Technology","ror":"https://ror.org/01zkghx44","country_code":"US","type":"education","lineage":["https://openalex.org/I130701444"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Saibal Mukhopadhyay","raw_affiliation_strings":["School of Electrical and Computer Engineering, Georgia Institute of Technology,Atlanta,Georgia,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Electrical and Computer Engineering, Georgia Institute of Technology,Atlanta,Georgia,USA","institution_ids":["https://openalex.org/I130701444"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I130701444"],"apc_list":null,"apc_paid":null,"fwci":1.117,"has_fulltext":false,"cited_by_count":2,"citation_normalized_percentile":{"value":0.74914819,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":95,"max":96},"biblio":{"volume":null,"issue":null,"first_page":"243","last_page":"247"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11181","display_name":"Advanced Data Storage Technologies","score":0.9995999932289124,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T11181","display_name":"Advanced Data Storage Technologies","score":0.9995999932289124,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T11269","display_name":"Algorithms and Data Compression","score":0.9976000189781189,"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/T11478","display_name":"Caching and Content Delivery","score":0.9948999881744385,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/cache","display_name":"Cache","score":0.7709469795227051},{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.6814991235733032},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.5979341864585876},{"id":"https://openalex.org/keywords/hessian-matrix","display_name":"Hessian matrix","score":0.4743279814720154},{"id":"https://openalex.org/keywords/parallel-computing","display_name":"Parallel computing","score":0.3254334330558777},{"id":"https://openalex.org/keywords/algorithm","display_name":"Algorithm","score":0.10849115252494812},{"id":"https://openalex.org/keywords/mathematics","display_name":"Mathematics","score":0.10564836859703064}],"concepts":[{"id":"https://openalex.org/C115537543","wikidata":"https://www.wikidata.org/wiki/Q165596","display_name":"Cache","level":2,"score":0.7709469795227051},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.6814991235733032},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5979341864585876},{"id":"https://openalex.org/C203616005","wikidata":"https://www.wikidata.org/wiki/Q620495","display_name":"Hessian matrix","level":2,"score":0.4743279814720154},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.3254334330558777},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.10849115252494812},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.10564836859703064},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.0}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/mwscas60917.2024.10658840","is_oa":false,"landing_page_url":"https://doi.org/10.1109/mwscas60917.2024.10658840","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2024 IEEE 67th International Midwest Symposium on Circuits and Systems (MWSCAS)","raw_type":"proceedings-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":28,"referenced_works":["https://openalex.org/W2896457183","https://openalex.org/W3110757807","https://openalex.org/W3161584439","https://openalex.org/W4292779060","https://openalex.org/W4322718191","https://openalex.org/W4384918448","https://openalex.org/W4386978016","https://openalex.org/W4388093177","https://openalex.org/W4390690436","https://openalex.org/W4391410261","https://openalex.org/W4391463154","https://openalex.org/W4391591383","https://openalex.org/W4402349265","https://openalex.org/W4402671766","https://openalex.org/W6778883912","https://openalex.org/W6784041521","https://openalex.org/W6795766024","https://openalex.org/W6846659131","https://openalex.org/W6850625674","https://openalex.org/W6850927664","https://openalex.org/W6854866820","https://openalex.org/W6855916431","https://openalex.org/W6856685690","https://openalex.org/W6857799723","https://openalex.org/W6860936625","https://openalex.org/W6861013483","https://openalex.org/W6862025885","https://openalex.org/W6893640197"],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2748952813","https://openalex.org/W2611031068","https://openalex.org/W1704347466","https://openalex.org/W4283017538","https://openalex.org/W1996936972","https://openalex.org/W1545275724","https://openalex.org/W2802707792","https://openalex.org/W2569979269","https://openalex.org/W2075777916"],"abstract_inverted_index":{"As":[0],"large":[1],"language":[2,16],"models":[3],"continue":[4],"to":[5,64,69,76,110,119],"advance,":[6],"handling":[7],"extended":[8],"context":[9],"lengths":[10],"becomes":[11],"crucial":[12],"for":[13,44],"improving":[14],"natural":[15],"processing":[17],"tasks.":[18],"However,":[19],"it":[20],"also":[21],"introduces":[22],"significant":[23,102],"computational":[24,145],"challenges,":[25],"particularly":[26],"in":[27,116,143,147],"managing":[28],"KV":[29,40,103],"cache":[30,41,104],"memory.":[31],"To":[32],"address":[33],"this,":[34],"we":[35],"present":[36],"a":[37,49,55,126],"novel":[38],"Hessian-aware":[39],"quantization":[42,58],"method":[43,47,100],"LLMs.":[45,148],"This":[46,123],"employs":[48],"weight-aware":[50],"head-wise":[51],"Hessian":[52],"analysis":[53],"and":[54,73,86,95],"three-tiered":[56],"mixed-precision":[57],"strategy.":[59],"It":[60],"assigns":[61],"high":[62],"precision":[63,68,75],"magnitude-based":[65],"outliers,":[66],"intermediate":[67],"Hessian-based":[70],"important":[71],"elements,":[72],"low":[74],"the":[77,120,133,137],"rest,":[78],"achieving":[79],"an":[80],"optimal":[81],"balance":[82],"between":[83],"memory":[84],"compression":[85,105],"accuracy.":[87],"Our":[88],"extensive":[89],"testing":[90],"across":[91],"various":[92],"LLM":[93],"configurations":[94],"datasets":[96],"shows":[97],"that":[98],"this":[99],"achieves":[101],"rates,":[106],"ranging":[107],"from":[108],"8.21x":[109],"9.7x":[111],"while":[112],"maintaining":[113],"minimal":[114],"deviation":[115],"accuracy":[117],"compared":[118],"FP16":[121,134],"baseline.":[122],"enhancement":[124],"enables":[125],"remarkable":[127],"3.8x":[128],"higher":[129],"throughput":[130],"(tokens/second)":[131],"than":[132],"baseline,":[135],"highlighting":[136],"practical":[138],"utility":[139],"of":[140],"our":[141],"approach":[142],"addressing":[144],"challenges":[146]},"counts_by_year":[{"year":2025,"cited_by_count":2}],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2025-10-10T00:00:00"}