{"id":"https://openalex.org/W7140313435","doi":"https://doi.org/10.1109/nvmsa66678.2025.00014","title":"Memory-Optimized Offloading: Enabling LLM Fine-Tuning on Memory-Constrained Hardware","display_name":"Memory-Optimized Offloading: Enabling LLM Fine-Tuning on Memory-Constrained Hardware","publication_year":2025,"publication_date":"2025-08-20","ids":{"openalex":"https://openalex.org/W7140313435","doi":"https://doi.org/10.1109/nvmsa66678.2025.00014"},"language":null,"primary_location":{"id":"doi:10.1109/nvmsa66678.2025.00014","is_oa":false,"landing_page_url":"https://doi.org/10.1109/nvmsa66678.2025.00014","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE 14th Non-Volatile Memory Systems and Applications Symposium (NVMSA)","raw_type":"proceedings-article"},"type":"article","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/A5108213965","display_name":"Kaihuan Lin","orcid":null},"institutions":[{"id":"https://openalex.org/I191208505","display_name":"Xiamen University","ror":"https://ror.org/00mcjh785","country_code":"CN","type":"education","lineage":["https://openalex.org/I191208505"]}],"countries":["CN"],"is_corresponding":true,"raw_author_name":"Kaihuan Lin","raw_affiliation_strings":["Xiamen University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Xiamen University","institution_ids":["https://openalex.org/I191208505"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5091044785","display_name":"Hongchao Du","orcid":"https://orcid.org/0000-0001-7995-1834"},"institutions":[{"id":"https://openalex.org/I168719708","display_name":"City University of Hong Kong","ror":"https://ror.org/03q8dnn23","country_code":"HK","type":"education","lineage":["https://openalex.org/I168719708"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Hongchao Du","raw_affiliation_strings":["City University of Hong Kong"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"City University of Hong Kong","institution_ids":["https://openalex.org/I168719708"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130596492","display_name":"Dawei Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dawei Wu","raw_affiliation_strings":["YEESTOR Microelectronics Co., Ltd"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"YEESTOR Microelectronics Co., Ltd","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130561801","display_name":"Yin Lin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yin Lin","raw_affiliation_strings":["YEESTOR Microelectronics Co., Ltd"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"YEESTOR Microelectronics Co., Ltd","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130608791","display_name":"Ting Luo","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ting Luo","raw_affiliation_strings":["YEESTOR Microelectronics Co., Ltd"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"YEESTOR Microelectronics Co., Ltd","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100393695","display_name":"Qiao Li","orcid":"https://orcid.org/0000-0002-4579-4268"},"institutions":[{"id":"https://openalex.org/I4210113480","display_name":"Mohamed bin Zayed University of Artificial Intelligence","ror":"https://ror.org/0258gkt32","country_code":"AE","type":"education","lineage":["https://openalex.org/I4210113480"]}],"countries":["AE"],"is_corresponding":false,"raw_author_name":"Qiao Li","raw_affiliation_strings":["Mohamed bin Zayed University of Artificial Intelligence"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Mohamed bin Zayed University of Artificial Intelligence","institution_ids":["https://openalex.org/I4210113480"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5127815307","display_name":"Chun Jason Xue","orcid":null},"institutions":[{"id":"https://openalex.org/I4210113480","display_name":"Mohamed bin Zayed University of Artificial Intelligence","ror":"https://ror.org/0258gkt32","country_code":"AE","type":"education","lineage":["https://openalex.org/I4210113480"]}],"countries":["AE"],"is_corresponding":false,"raw_author_name":"Chun Jason Xue","raw_affiliation_strings":["Mohamed bin Zayed University of Artificial Intelligence"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Mohamed bin Zayed University of Artificial Intelligence","institution_ids":["https://openalex.org/I4210113480"]}]}],"institutions":[],"countries_distinct_count":3,"institutions_distinct_count":7,"corresponding_author_ids":["https://openalex.org/A5108213965"],"corresponding_institution_ids":["https://openalex.org/I191208505"],"apc_list":null,"apc_paid":null,"fwci":1.8963,"has_fulltext":false,"cited_by_count":1,"citation_normalized_percentile":{"value":0.90158924,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":95,"max":98},"biblio":{"volume":null,"issue":null,"first_page":"36","last_page":"41"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.6158000230789185,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.6158000230789185,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10904","display_name":"Embedded Systems Design Techniques","score":0.13300000131130219,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T12326","display_name":"Network Packet Processing and Optimization","score":0.02500000037252903,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/software","display_name":"Software","score":0.3337000012397766},{"id":"https://openalex.org/keywords/troubleshooting","display_name":"Troubleshooting","score":0.3059000074863434},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.30059999227523804},{"id":"https://openalex.org/keywords/pipeline","display_name":"Pipeline (software)","score":0.27619999647140503},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.26350000500679016}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5343000292778015},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.37209999561309814},{"id":"https://openalex.org/C9390403","wikidata":"https://www.wikidata.org/wiki/Q3966","display_name":"Computer hardware","level":1,"score":0.3508000075817108},{"id":"https://openalex.org/C2777904410","wikidata":"https://www.wikidata.org/wiki/Q7397","display_name":"Software","level":2,"score":0.3337000012397766},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.3149000108242035},{"id":"https://openalex.org/C147494362","wikidata":"https://www.wikidata.org/wiki/Q2078905","display_name":"Troubleshooting","level":2,"score":0.3059000074863434},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.30059999227523804},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.27619999647140503},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.26350000500679016},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.25189998745918274}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/nvmsa66678.2025.00014","is_oa":false,"landing_page_url":"https://doi.org/10.1109/nvmsa66678.2025.00014","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE 14th Non-Volatile Memory Systems and Applications Symposium (NVMSA)","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":13,"referenced_works":["https://openalex.org/W2339765813","https://openalex.org/W2489529491","https://openalex.org/W3012479151","https://openalex.org/W3012514909","https://openalex.org/W3035190488","https://openalex.org/W3081168214","https://openalex.org/W3129831491","https://openalex.org/W3149755221","https://openalex.org/W3189259198","https://openalex.org/W3205803342","https://openalex.org/W4378800959","https://openalex.org/W7133185942","https://openalex.org/W7140128959"],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1,7],"models":[2,25,83],"(LLMs)":[3],"have":[4],"revolutionized":[5],"natural":[6],"processing,":[8],"enabling":[9],"breakthroughs":[10],"in":[11],"applications":[12],"ranging":[13],"from":[14,64],"machine":[15],"translation":[16],"to":[17,37,144,193,214],"conversational":[18],"artificial":[19],"intelligence.":[20],"However,":[21],"fine-tuning":[22,32,49,80,100,109,202,225],"these":[23,61,89],"massive":[24],"presents":[26],"significant":[27],"memory":[28,57,74,105,134,152,186,213],"challenges,":[29,90],"especially":[30],"when":[31,148],"with":[33,177],"mixed":[34],"precision":[35],"(FP16/FP32)":[36],"maintain":[38],"efficiency":[39],"and":[40,68,111,146],"accuracy.":[41],"Although":[42],"existing":[43],"techniques,":[44],"such":[45],"as":[46],"the":[47,53,79,140,158,163,182,194,205],"DeepSpeed":[48,195],"framework,":[50],"can":[51,208],"reduce":[52],"overhead":[54,106],"of":[55,81,129,162,211],"host":[56],"using":[58],"group":[59],"fine-tuning,":[60],"methods":[62],"suffer":[63],"inefficient":[65],"overflow":[66,121],"detection":[67],"redundant":[69],"gradient":[70,96,138,141,156,164],"storage,":[71],"greatly":[72],"increasing":[73],"requirements.":[75],"These":[76,220],"drawbacks":[77],"limit":[78],"large":[82],"on":[84,227],"memory-constrained":[85],"devices.":[86],"To":[87],"address":[88],"this":[91],"paper":[92],"proposes":[93],"an":[94],"instant":[95],"offloading":[97],"strategy,":[98],"a":[99,200,216],"approach":[101],"that":[102,169,181],"significantly":[103],"reduces":[104,185],"while":[107,198],"maintaining":[108,199],"accuracy":[110],"speed.":[112,203],"Our":[113],"solution":[114],"introduces":[115],"three":[116],"key":[117],"innovations:":[118],"(1)":[119],"On-the-fly":[120],"detection:":[122],"calculated":[123],"gradients":[124],"are":[125],"checked":[126],"immediately":[127],"instead":[128],"batch":[130],"checking,":[131],"eliminating":[132],"peak":[133],"spikes;":[135],"(2)":[136],"Instant":[137],"offloading:":[139,157],"is":[142,166],"offloaded":[143],"SSD":[145],"loaded":[147],"needed,":[149],"minimizing":[150],"CPU":[151,212],"usage;":[153],"(3)":[154],"FP16-centric":[155],"type":[159],"conversion":[160],"operation":[161],"data":[165,175],"delayed":[167],"so":[168],"it":[170],"uses":[171],"FP16":[172],"format":[173],"for":[174],"exchange":[176],"SSD.":[178],"Experiments":[179],"show":[180],"proposed":[183,206],"design":[184],"consumption":[187],"by":[188],"77.8%":[189],"\u2212":[190],"85.3%":[191],"compared":[192],"offload":[196],"baseline":[197],"comparable":[201],"Specifically,":[204],"system":[207],"use":[209],"128GB":[210],"fine-tune":[215],"98B":[217],"parameter":[218],"model.":[219],"advances":[221],"make":[222],"large-scale":[223],"LLM":[224],"possible":[226],"resource-constrained":[228],"hardware.":[229]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-05-28T09:10:13.091523","created_date":"2026-03-26T00:00:00"}