{"id":"https://openalex.org/W4405172828","doi":"https://doi.org/10.48550/arxiv.2412.05270","title":"APOLLO: SGD-like Memory, AdamW-level Performance","display_name":"APOLLO: SGD-like Memory, AdamW-level Performance","publication_year":2024,"publication_date":"2024-12-06","ids":{"openalex":"https://openalex.org/W4405172828","doi":"https://doi.org/10.48550/arxiv.2412.05270"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2412.05270","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2412.05270","pdf_url":"https://arxiv.org/pdf/2412.05270","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,"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"},"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2412.05270","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100567749","display_name":"Hanqing Zhu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhu, Hanqing","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5115057044","display_name":"Zhenyu Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Zhenyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Cong, Wenyan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cong, Wenyan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5011242991","display_name":"Xijun Liu","orcid":"https://orcid.org/0000-0002-7219-5914"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Xi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Park, Sem","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Park, Sem","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5016704219","display_name":"Vikas Chandra","orcid":"https://orcid.org/0009-0005-4996-8455"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chandra, Vikas","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5003962662","display_name":"Bo Long","orcid":"https://orcid.org/0000-0002-8149-0317"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Long, Bo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5011883763","display_name":"David Z. Pan","orcid":"https://orcid.org/0000-0002-5705-2501"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pan, David Z.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5048522863","display_name":"Zhangyang Wang","orcid":"https://orcid.org/0000-0002-2050-5693"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Zhangyang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"Lee, Jinwon","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lee, Jinwon","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":0,"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.9348999857902527,"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":0.9348999857902527,"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/apollo","display_name":"Apollo","score":0.7797126173973083},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.4057326912879944},{"id":"https://openalex.org/keywords/biology","display_name":"Biology","score":0.15738192200660706},{"id":"https://openalex.org/keywords/zoology","display_name":"Zoology","score":0.09445589780807495}],"concepts":[{"id":"https://openalex.org/C2779821442","wikidata":"https://www.wikidata.org/wiki/Q41633","display_name":"Apollo","level":2,"score":0.7797126173973083},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4057326912879944},{"id":"https://openalex.org/C86803240","wikidata":"https://www.wikidata.org/wiki/Q420","display_name":"Biology","level":0,"score":0.15738192200660706},{"id":"https://openalex.org/C90856448","wikidata":"https://www.wikidata.org/wiki/Q431","display_name":"Zoology","level":1,"score":0.09445589780807495}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2412.05270","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2412.05270","pdf_url":"https://arxiv.org/pdf/2412.05270","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,"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.2412.05270","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2412.05270","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,"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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2412.05270","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2412.05270","pdf_url":"https://arxiv.org/pdf/2412.05270","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,"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":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4405172828.pdf","grobid_xml":"https://content.openalex.org/works/W4405172828.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2899084033","https://openalex.org/W2748952813","https://openalex.org/W2054178368","https://openalex.org/W2001637022","https://openalex.org/W2749643947","https://openalex.org/W2921637885","https://openalex.org/W4244081656","https://openalex.org/W2981806746","https://openalex.org/W4400809088"],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2],"(LLMs)":[3],"are":[4],"notoriously":[5],"memory-intensive":[6],"during":[7],"training,":[8],"particularly":[9],"with":[10,158,171,219,246],"the":[11,166,184],"popular":[12],"AdamW":[13,157,206],"optimizer.":[14],"This":[15,127],"memory":[16,45,70,139,160,179,245],"burden":[17],"necessitates":[18],"using":[19,116,239],"more":[20],"or":[21,24,172],"higher-end":[22],"GPUs":[23,224],"reducing":[25],"batch":[26,211],"sizes,":[27],"limiting":[28],"training":[29],"scalability":[30],"and":[31,65],"throughput.":[32],"To":[33],"address":[34],"this,":[35],"various":[36],"memory-efficient":[37],"optimizers":[38],"have":[39],"been":[40],"proposed":[41],"to":[42,63,72,137,156,205],"reduce":[43],"optimizer":[44,69,120],"usage.":[46],"However,":[47],"they":[48],"face":[49],"critical":[50],"challenges:":[51],"(i)":[52],"reliance":[53],"on":[54,98,123,200,222,235],"costly":[55],"SVD":[56],"operations;":[57],"(ii)":[58],"significant":[59,192],"performance":[60,154],"trade-offs":[61],"compared":[62,155,204],"AdamW;":[64],"(iii)":[66],"still":[67],"substantial":[68],"overhead":[71],"maintain":[73],"competitive":[74],"performance.":[75,145],"In":[76],"this":[77,99],"work,":[78],"we":[79,101],"identify":[80],"that":[81,165],"AdamW's":[82],"learning":[83,94,113,129],"rate":[84,95,114,130],"adaptation":[85],"rule":[86,132],"can":[87],"be":[88],"effectively":[89],"coarsened":[90],"as":[91],"a":[92,236],"structured":[93,128],"update.":[96],"Based":[97],"insight,":[100],"propose":[102],"Approximated":[103],"Gradient":[104],"Scaling":[105],"for":[106],"Memory-Efficient":[107],"LLM":[108],"Optimization":[109],"(APOLLO),":[110],"which":[111],"approximates":[112],"scaling":[115],"an":[117,201],"auxiliary":[118],"low-rank":[119],"state":[121],"based":[122],"pure":[124],"random":[125],"projection.":[126],"update":[131],"makes":[133],"APOLLO":[134,167],"highly":[135],"tolerant":[136],"further":[138],"reductions":[140],"while":[141,176],"delivering":[142],"comparable":[143],"pre-training":[144,153],"Even":[146],"its":[147],"rank-1":[148],"variant,":[149],"APOLLO-Mini,":[150],"achieves":[151],"superior":[152],"SGD-level":[159],"costs.":[161],"Extensive":[162],"experiments":[163],"demonstrate":[164],"series":[168],"performs":[169],"on-par":[170],"better":[173],"than":[174,241],"AdamW,":[175],"achieving":[177],"greater":[178],"savings":[180,190],"by":[181,207],"nearly":[182],"eliminating":[183],"optimization":[185],"states":[186],"of":[187,244],"AdamW.":[188],"These":[189],"provide":[191],"system-level":[193,226],"benefits:":[194],"(1)":[195],"Enhanced":[196],"Throughput:":[197],"3x":[198],"throughput":[199],"8xA100-80GB":[202],"setup":[203],"supporting":[208],"4x":[209],"larger":[210],"sizes.":[212],"(2)":[213],"Improved":[214],"Model":[215],"Scalability:":[216],"Pre-training":[217,233],"LLaMA-13B":[218],"naive":[220],"DDP":[221],"A100-80GB":[223],"without":[225],"optimizations.":[227],"(3)":[228],"Low-End":[229],"GPU":[230,238],"Friendly":[231],"Pre-training:":[232],"LLaMA-7B":[234],"single":[237],"less":[240],"12":[242],"GB":[243],"weight":[247],"quantization.":[248]},"counts_by_year":[],"updated_date":"2026-08-22T07:34:49.880490","created_date":"2025-10-10T00:00:00"}