{"id":"https://openalex.org/W7155172506","doi":"https://doi.org/10.48550/arxiv.2604.19503","title":"ReaLB: Real-Time Load Balancing for Multimodal MoE Inference","display_name":"ReaLB: Real-Time Load Balancing for Multimodal MoE Inference","publication_year":2026,"publication_date":"2026-04-21","ids":{"openalex":"https://openalex.org/W7155172506","doi":"https://doi.org/10.48550/arxiv.2604.19503"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.19503","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.19503","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":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2604.19503","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5115591843","display_name":"Yu Wang","orcid":"https://orcid.org/0009-0005-0661-311X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Yingping","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134342922","display_name":"Yi Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Yi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134332074","display_name":"Xiangyu Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Xiangyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134295228","display_name":"Junwei Cui","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cui, Junwei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5049735535","display_name":"\u8521\u73ae\u7433","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cai, Weilin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134253452","display_name":"Zhijiang Guo","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guo, Zhijiang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5134352994","display_name":"Jiayi Huang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huang, Jiayi","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":false,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.16820000112056732,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.16820000112056732,"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":0.11429999768733978,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.10970000177621841,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.6484000086784363},{"id":"https://openalex.org/keywords/speedup","display_name":"Speedup","score":0.6449000239372253},{"id":"https://openalex.org/keywords/computation","display_name":"Computation","score":0.612500011920929},{"id":"https://openalex.org/keywords/scheduling","display_name":"Scheduling (production processes)","score":0.5539000034332275},{"id":"https://openalex.org/keywords/load-balancing","display_name":"Load balancing (electrical power)","score":0.541700005531311},{"id":"https://openalex.org/keywords/overhead","display_name":"Overhead (engineering)","score":0.3880999982357025},{"id":"https://openalex.org/keywords/limiting","display_name":"Limiting","score":0.3718000054359436},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.35120001435279846}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8356999754905701},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6484000086784363},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.6449000239372253},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.612500011920929},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.5539000034332275},{"id":"https://openalex.org/C138959212","wikidata":"https://www.wikidata.org/wiki/Q1806783","display_name":"Load balancing (electrical power)","level":3,"score":0.541700005531311},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.3880999982357025},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.3718000054359436},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.35120001435279846},{"id":"https://openalex.org/C58328972","wikidata":"https://www.wikidata.org/wiki/Q184609","display_name":"Expert system","level":2,"score":0.35010001063346863},{"id":"https://openalex.org/C2781020372","wikidata":"https://www.wikidata.org/wiki/Q533093","display_name":"On the fly","level":2,"score":0.34470000863075256},{"id":"https://openalex.org/C2777904410","wikidata":"https://www.wikidata.org/wiki/Q7397","display_name":"Software","level":2,"score":0.31610000133514404},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.30219998955726624},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2976999878883362},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.2883000075817108},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.2766000032424927},{"id":"https://openalex.org/C2989134064","wikidata":"https://www.wikidata.org/wiki/Q288510","display_name":"Execution time","level":2,"score":0.2745000123977661},{"id":"https://openalex.org/C2779370713","wikidata":"https://www.wikidata.org/wiki/Q357554","display_name":"Load management","level":2,"score":0.27250000834465027},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.2662000060081482},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2653000056743622},{"id":"https://openalex.org/C46743427","wikidata":"https://www.wikidata.org/wiki/Q1341685","display_name":"Inference engine","level":3,"score":0.262800008058548},{"id":"https://openalex.org/C2777131613","wikidata":"https://www.wikidata.org/wiki/Q18394147","display_name":"Retard","level":2,"score":0.2556000053882599},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.2531000077724457},{"id":"https://openalex.org/C107568181","wikidata":"https://www.wikidata.org/wiki/Q5319000","display_name":"Dynamic priority scheduling","level":3,"score":0.2515000104904175}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.19503","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.19503","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":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2604.19503","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.19503","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":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Mixture-of-Experts":[0],"(MoE)":[1],"architectures":[2],"are":[3],"widely":[4],"used":[5],"in":[6],"modern":[7],"large":[8,35],"language":[9],"models":[10,154],"and":[11,23,138],"multimodal":[12,75],"models.":[13],"However,":[14],"inference":[15,78],"efficiency":[16,112],"is":[17],"often":[18],"limited":[19],"by":[20,102,113],"highly":[21],"dynamic":[22],"skewed":[24],"expert":[25,46,132],"workloads":[26],"across":[27],"different":[28],"modalities.":[29],"During":[30],"the":[31,42,87,136,140,144],"prefill":[32],"stage":[33],"with":[34],"batch":[36],"sizes,":[37],"vision":[38],"tokens":[39],"frequently":[40],"dominate":[41],"input":[43],"sequences.":[44],"Under":[45],"parallelism":[47],"(EP),":[48],"this":[49],"leads":[50],"to":[51,109,167],"severe":[52],"load":[53,71],"imbalance,":[54],"where":[55],"a":[56,69,96],"subset":[57],"of":[58,90],"devices":[59],"becomes":[60],"overloaded,":[61],"reducing":[62],"overall":[63],"system":[64],"throughput.":[65],"We":[66],"propose":[67],"ReaLB,":[68],"real-time":[70],"balancing":[72],"method":[73],"for":[74],"MoE":[76,91,148],"(MMoE)":[77],"that":[79,156],"introduces":[80],"zero":[81],"scheduling":[82],"overhead.":[83],"ReaLB":[84,105,118,157],"dynamically":[85],"adjusts":[86],"computation":[88,108],"precision":[89,133],"experts":[92,123],"at":[93],"runtime":[94],"on":[95,135,151],"per-EP-rank":[97],"basis.":[98],"For":[99],"ranks":[100],"dominated":[101],"vision-heavy":[103],"experts,":[104],"assigns":[106],"lower-precision":[107],"improve":[110],"execution":[111],"exploiting":[114],"FP4":[115],"Tensor":[116],"Cores.":[117],"does":[119],"not":[120],"require":[121],"redundant":[122],"or":[124],"additional":[125],"memory":[126],"allocation.":[127],"Instead,":[128],"it":[129],"performs":[130],"layer-wise":[131],"transformation":[134],"fly":[137],"hides":[139],"associated":[141],"overhead":[142],"within":[143,168],"dispatch":[145],"phase":[146],"before":[147],"computation.":[149],"Experiments":[150],"representative":[152],"MMoE":[153],"show":[155],"achieves":[158],"1.10$\\times$-1.32$\\times$":[159],"end-to-end":[160],"speedup":[161],"while":[162],"limiting":[163],"average":[164],"accuracy":[165],"degradation":[166],"1%.":[169]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-04-23T00:00:00"}