{"id":"https://openalex.org/W4416123554","doi":"https://doi.org/10.48550/arxiv.2504.02263","title":"MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism","display_name":"MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism","publication_year":2025,"publication_date":"2025-04-03","ids":{"openalex":"https://openalex.org/W4416123554","doi":"https://doi.org/10.48550/arxiv.2504.02263"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2504.02263","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2504.02263","pdf_url":"https://arxiv.org/pdf/2504.02263","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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/2504.02263","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5101852647","display_name":"Ruidong Zhu","orcid":"https://orcid.org/0000-0003-1389-7884"},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Zhu, Ruidong","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5066345089","display_name":"Ziheng Jiang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiang, Ziheng","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5074344378","display_name":"Chao Jin","orcid":"https://orcid.org/0000-0001-6590-333X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jin, Chao","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5108630261","display_name":"Peng Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Peng","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5068036753","display_name":"Cesar A. Stuardo","orcid":"https://orcid.org/0000-0002-5052-0077"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Stuardo, Cesar A.","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5113142890","display_name":"Dongyang Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Dongyang","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5031506917","display_name":"Xinlei Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Xinlei","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5031853666","display_name":"Huaping Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Huaping","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5033966516","display_name":"Haoran Wei","orcid":"https://orcid.org/0009-0004-4093-3187"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wei, Haoran","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5060417049","display_name":"Cheng Yang","orcid":"https://orcid.org/0000-0001-7821-0030"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cheng, Yang","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Xiao, Jianzhe","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiao, Jianzhe","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Zhang, Xinyi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Xinyi","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5007420841","display_name":"L. Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Lingjun","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5084047386","display_name":"Haibin Lin","orcid":"https://orcid.org/0000-0003-4879-5335"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lin, Haibin","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5043931807","display_name":"Li\u2010Wen Chang","orcid":"https://orcid.org/0000-0001-6515-6733"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chang, Li-Wen","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5007239185","display_name":"Jianxi Ye","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ye, Jianxi","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100547884","display_name":"Yu Xiao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Xiao","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5052249316","display_name":"Xuanzhe Liu","orcid":"https://orcid.org/0000-0002-7908-8484"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Xuanzhe","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101882004","display_name":"Xin Jin","orcid":"https://orcid.org/0000-0001-8741-5847"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jin, Xin","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5113343221","display_name":"Xin Liu","orcid":"https://orcid.org/0009-0000-8346-3323"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Xin","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":20,"corresponding_author_ids":["https://openalex.org/A5101852647"],"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/T10036","display_name":"Advanced Neural Network Applications","score":0.38769999146461487,"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":0.38769999146461487,"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/T14347","display_name":"Big Data and Digital Economy","score":0.07169999927282333,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.06539999693632126,"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/exploit","display_name":"Exploit","score":0.782800018787384},{"id":"https://openalex.org/keywords/pipeline","display_name":"Pipeline (software)","score":0.7202000021934509},{"id":"https://openalex.org/keywords/overhead","display_name":"Overhead (engineering)","score":0.7134000062942505},{"id":"https://openalex.org/keywords/initialization","display_name":"Initialization","score":0.6123999953269958},{"id":"https://openalex.org/keywords/parallelism","display_name":"Parallelism (grammar)","score":0.545199990272522},{"id":"https://openalex.org/keywords/throughput","display_name":"Throughput","score":0.5370000004768372},{"id":"https://openalex.org/keywords/scale","display_name":"Scale (ratio)","score":0.4830999970436096}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8596000075340271},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.782800018787384},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.7202000021934509},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.7134000062942505},{"id":"https://openalex.org/C114466953","wikidata":"https://www.wikidata.org/wiki/Q6034165","display_name":"Initialization","level":2,"score":0.6123999953269958},{"id":"https://openalex.org/C2781172179","wikidata":"https://www.wikidata.org/wiki/Q853109","display_name":"Parallelism (grammar)","level":2,"score":0.545199990272522},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.5370000004768372},{"id":"https://openalex.org/C2778755073","wikidata":"https://www.wikidata.org/wiki/Q10858537","display_name":"Scale (ratio)","level":2,"score":0.4830999970436096},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.48240000009536743},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.4345000088214874},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.4018000066280365},{"id":"https://openalex.org/C761482","wikidata":"https://www.wikidata.org/wiki/Q118093","display_name":"Transmission (telecommunications)","level":2,"score":0.3837999999523163},{"id":"https://openalex.org/C61483411","wikidata":"https://www.wikidata.org/wiki/Q3124522","display_name":"Data parallelism","level":3,"score":0.33649998903274536},{"id":"https://openalex.org/C557945733","wikidata":"https://www.wikidata.org/wiki/Q389772","display_name":"Data transmission","level":2,"score":0.3188999891281128},{"id":"https://openalex.org/C31258907","wikidata":"https://www.wikidata.org/wiki/Q1301371","display_name":"Computer network","level":1,"score":0.30250000953674316},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.28850001096725464},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.27320000529289246},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.2653999924659729},{"id":"https://openalex.org/C34165917","wikidata":"https://www.wikidata.org/wiki/Q188267","display_name":"Programming paradigm","level":2,"score":0.26019999384880066},{"id":"https://openalex.org/C2778348673","wikidata":"https://www.wikidata.org/wiki/Q739302","display_name":"Production (economics)","level":2,"score":0.2574000060558319}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2504.02263","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2504.02263","pdf_url":"https://arxiv.org/pdf/2504.02263","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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.2504.02263","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2504.02263","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":"article"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2504.02263","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2504.02263","pdf_url":"https://arxiv.org/pdf/2504.02263","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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/W4416123554.pdf","grobid_xml":"https://content.openalex.org/works/W4416123554.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Mixture-of-Experts":[0],"(MoE)":[1],"showcases":[2],"tremendous":[3],"potential":[4],"to":[5,30,35,129,169],"scale":[6],"large":[7],"language":[8],"models":[9],"(LLMs)":[10],"with":[11,111],"enhanced":[12],"performance":[13],"and":[14,40,49,60,73,101,106,123,132,135,159],"reduced":[15],"computational":[16],"complexity.":[17],"However,":[18],"its":[19],"sparsely":[20],"activated":[21],"architecture":[22],"shifts":[23],"feed-forward":[24],"networks":[25],"(FFNs)":[26],"from":[27],"being":[28],"compute-intensive":[29],"memory-intensive":[31],"during":[32],"inference,":[33],"leading":[34],"substantially":[36],"lower":[37],"GPU":[38,125,160],"utilization":[39],"increased":[41],"operational":[42],"costs.":[43],"We":[44],"present":[45],"MegaScale-Infer,":[46],"an":[47],"efficient":[48],"cost-effective":[50],"system":[51],"for":[52,76,108,115],"serving":[53],"large-scale":[54],"MoE":[55],"models.":[56],"MegaScale-Infer":[57,89,118,143,166],"disaggregates":[58],"attention":[59,105,131],"FFN":[61,133],"modules":[62,134],"within":[63],"each":[64,116],"model":[65,113],"layer,":[66],"enabling":[67],"independent":[68],"scaling,":[69],"tailored":[70],"parallelism":[71,114],"strategies,":[72],"heterogeneous":[74],"deployment":[75],"both":[77],"modules.":[78],"To":[79,127],"fully":[80],"exploit":[81],"disaggregation":[82],"in":[83],"the":[84],"presence":[85],"of":[86],"MoE's":[87],"sparsity,":[88],"introduces":[90],"ping-pong":[91],"pipeline":[92],"parallelism,":[93],"which":[94],"partitions":[95],"a":[96,145],"request":[97],"batch":[98],"into":[99],"micro-batches":[100],"shuttles":[102],"them":[103],"between":[104],"FFNs":[107],"inference.":[109],"Combined":[110],"distinct":[112],"module,":[117],"effectively":[119],"hides":[120],"communication":[121,148],"overhead":[122,139],"maximizes":[124],"utilization.":[126],"adapt":[128],"disaggregated":[130],"minimize":[136],"data":[137,154],"transmission":[138],"(e.g.,":[140],"token":[141],"dispatch),":[142],"provides":[144],"high-performance":[146],"M2N":[147],"library":[149],"that":[150,165],"eliminates":[151],"unnecessary":[152],"GPU-to-CPU":[153],"copies,":[155],"group":[156],"initialization":[157],"overhead,":[158],"synchronization.":[161],"Experimental":[162],"results":[163],"indicate":[164],"achieves":[167],"up":[168],"1.90x":[170],"higher":[171],"per-GPU":[172],"throughput":[173],"than":[174],"state-of-the-art":[175],"solutions.":[176]},"counts_by_year":[],"updated_date":"2026-04-17T18:11:37.981687","created_date":"2025-10-10T00:00:00"}