{"id":"https://openalex.org/W4409047962","doi":"https://doi.org/10.1145/3721146.3721940","title":"Accelerating MoE Model Inference with Expert Sharding","display_name":"Accelerating MoE Model Inference with Expert Sharding","publication_year":2025,"publication_date":"2025-03-30","ids":{"openalex":"https://openalex.org/W4409047962","doi":"https://doi.org/10.1145/3721146.3721940"},"language":"en","primary_location":{"id":"doi:10.1145/3721146.3721940","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3721146.3721940","pdf_url":"https://dl.acm.org/doi/pdf/10.1145/3721146.3721940","source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 5th Workshop on Machine Learning and Systems","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://dl.acm.org/doi/pdf/10.1145/3721146.3721940","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5055620336","display_name":"Oana Balmau","orcid":"https://orcid.org/0000-0002-6822-8891"},"institutions":[{"id":"https://openalex.org/I5023651","display_name":"McGill University","ror":"https://ror.org/01pxwe438","country_code":"CA","type":"education","lineage":["https://openalex.org/I5023651"]}],"countries":["CA"],"is_corresponding":true,"raw_author_name":"Oana Balmau","raw_affiliation_strings":["McGill University, Montreal, Canada"],"raw_orcid":"https://orcid.org/0000-0002-6822-8891","affiliations":[{"raw_affiliation_string":"McGill University, Montreal, Canada","institution_ids":["https://openalex.org/I5023651"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5031849486","display_name":"Anne-Marie Kermarrec","orcid":"https://orcid.org/0000-0001-8187-724X"},"institutions":[{"id":"https://openalex.org/I5124864","display_name":"\u00c9cole Polytechnique F\u00e9d\u00e9rale de Lausanne","ror":"https://ror.org/02s376052","country_code":"CH","type":"education","lineage":["https://openalex.org/I2799323385","https://openalex.org/I5124864"]}],"countries":["CH"],"is_corresponding":false,"raw_author_name":"Anne-Marie Kermarrec","raw_affiliation_strings":["EPFL, Lausanne, Switzerland"],"raw_orcid":"https://orcid.org/0000-0001-8187-724X","affiliations":[{"raw_affiliation_string":"EPFL, Lausanne, Switzerland","institution_ids":["https://openalex.org/I5124864"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5000635510","display_name":"Rafael Pires","orcid":"https://orcid.org/0000-0002-7826-1599"},"institutions":[{"id":"https://openalex.org/I5124864","display_name":"\u00c9cole Polytechnique F\u00e9d\u00e9rale de Lausanne","ror":"https://ror.org/02s376052","country_code":"CH","type":"education","lineage":["https://openalex.org/I2799323385","https://openalex.org/I5124864"]}],"countries":["CH"],"is_corresponding":false,"raw_author_name":"Rafael Pires","raw_affiliation_strings":["EPFL, Lausanne, Switzerland"],"raw_orcid":"https://orcid.org/0000-0002-7826-1599","affiliations":[{"raw_affiliation_string":"EPFL, Lausanne, Switzerland","institution_ids":["https://openalex.org/I5124864"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5116867205","display_name":"Andr\u00e9 Loureiro Esp\u00edrito Santo","orcid":null},"institutions":[{"id":"https://openalex.org/I5124864","display_name":"\u00c9cole Polytechnique F\u00e9d\u00e9rale de Lausanne","ror":"https://ror.org/02s376052","country_code":"CH","type":"education","lineage":["https://openalex.org/I2799323385","https://openalex.org/I5124864"]}],"countries":["CH"],"is_corresponding":false,"raw_author_name":"Andr\u00e9 Loureiro Esp\u00edrito Santo","raw_affiliation_strings":["EPFL, Lausanne, Switzerland"],"raw_orcid":"https://orcid.org/0009-0006-0111-1863","affiliations":[{"raw_affiliation_string":"EPFL, Lausanne, Switzerland","institution_ids":["https://openalex.org/I5124864"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5010233454","display_name":"Martijn de Vos","orcid":"https://orcid.org/0000-0003-4157-4847"},"institutions":[{"id":"https://openalex.org/I5124864","display_name":"\u00c9cole Polytechnique F\u00e9d\u00e9rale de Lausanne","ror":"https://ror.org/02s376052","country_code":"CH","type":"education","lineage":["https://openalex.org/I2799323385","https://openalex.org/I5124864"]}],"countries":["CH"],"is_corresponding":false,"raw_author_name":"Martijn de Vos","raw_affiliation_strings":["EPFL, Lausanne, Switzerland"],"raw_orcid":"https://orcid.org/0000-0003-4157-4847","affiliations":[{"raw_affiliation_string":"EPFL, Lausanne, Switzerland","institution_ids":["https://openalex.org/I5124864"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5047520340","display_name":"Milos Vujasinovic","orcid":"https://orcid.org/0009-0006-5083-3045"},"institutions":[{"id":"https://openalex.org/I5124864","display_name":"\u00c9cole Polytechnique F\u00e9d\u00e9rale de Lausanne","ror":"https://ror.org/02s376052","country_code":"CH","type":"education","lineage":["https://openalex.org/I2799323385","https://openalex.org/I5124864"]}],"countries":["CH"],"is_corresponding":false,"raw_author_name":"Milos Vujasinovic","raw_affiliation_strings":["EPFL, Lausanne, Switzerland"],"raw_orcid":"https://orcid.org/0009-0006-5083-3045","affiliations":[{"raw_affiliation_string":"EPFL, Lausanne, Switzerland","institution_ids":["https://openalex.org/I5124864"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":6,"corresponding_author_ids":["https://openalex.org/A5055620336"],"corresponding_institution_ids":["https://openalex.org/I5023651"],"apc_list":null,"apc_paid":null,"fwci":4.1552,"has_fulltext":false,"cited_by_count":2,"citation_normalized_percentile":{"value":0.93223695,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":98,"max":99},"biblio":{"volume":null,"issue":null,"first_page":"192","last_page":"199"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.9995999932289124,"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"}},"topics":[{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.9995999932289124,"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/T13702","display_name":"Machine Learning in Healthcare","score":0.9958999752998352,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9846000075340271,"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/inference","display_name":"Inference","score":0.6990297436714172},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.6466267108917236},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.348802387714386}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6990297436714172},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6466267108917236},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.348802387714386}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1145/3721146.3721940","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3721146.3721940","pdf_url":"https://dl.acm.org/doi/pdf/10.1145/3721146.3721940","source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 5th Workshop on Machine Learning and Systems","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.1145/3721146.3721940","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3721146.3721940","pdf_url":"https://dl.acm.org/doi/pdf/10.1145/3721146.3721940","source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 5th Workshop on Machine Learning and Systems","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4409047962.pdf"},"referenced_works_count":26,"referenced_works":["https://openalex.org/W1566289585","https://openalex.org/W2581624817","https://openalex.org/W2734941459","https://openalex.org/W2926767350","https://openalex.org/W2981852735","https://openalex.org/W2990138404","https://openalex.org/W3040573126","https://openalex.org/W3101708369","https://openalex.org/W3138895808","https://openalex.org/W3192523796","https://openalex.org/W4224308101","https://openalex.org/W4226515448","https://openalex.org/W4281922990","https://openalex.org/W4287748830","https://openalex.org/W4293718192","https://openalex.org/W4310510250","https://openalex.org/W4321610465","https://openalex.org/W4376272906","https://openalex.org/W4376652719","https://openalex.org/W4388855670","https://openalex.org/W4391012448","https://openalex.org/W4400409890","https://openalex.org/W6793032698","https://openalex.org/W6800054401","https://openalex.org/W6810737565","https://openalex.org/W6839827798"],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2899084033","https://openalex.org/W2748952813","https://openalex.org/W2390279801","https://openalex.org/W4391913857","https://openalex.org/W2358668433","https://openalex.org/W4396701345","https://openalex.org/W2376932109","https://openalex.org/W2001405890","https://openalex.org/W4396696052"],"abstract_inverted_index":{"Mixture":[0],"of":[1,65,94,107,145],"experts":[2],"(MoE)":[3],"models":[4,41],"achieve":[5,98],"state-of-the-art":[6],"results":[7,156],"in":[8,42,149],"language":[9],"modeling":[10],"but":[11],"suffer":[12],"from":[13],"inefficient":[14],"hardware":[15],"utilization":[16,92],"due":[17],"to":[18,147,151,162],"imbalanced":[19,119],"token":[20,89,153],"routing":[21,95],"and":[22,34,86,104,114,169],"communication":[23],"overhead.":[24],"While":[25],"prior":[26],"work":[27],"has":[28],"focused":[29],"on":[30,73,140],"optimizing":[31],"MoE":[32,40,66,174],"training":[33],"decoder":[35],"architectures,":[36,142],"inference":[37,55],"for":[38,172],"encoder-based":[39,141],"a":[43,101,167],"multi-GPU":[44],"with":[45],"expert":[46,108,120,130],"parallelism":[47],"setting":[48],"remains":[49],"underexplored.":[50],"We":[51,97,135],"introduce":[52],"MoEShard,":[53],"an":[54],"system":[56],"that":[57,71,158],"achieves":[58],"perfect":[59],"load":[60],"balancing":[61],"through":[62,100],"tensor":[63,164],"sharding":[64,165],"experts.":[67],"Unlike":[68],"existing":[69],"approaches":[70],"rely":[72],"heuristic":[74],"capacity":[75],"factors":[76],"or":[77],"drop":[78],"tokens,":[79],"MoEShard":[80,123,137],"evenly":[81],"distributes":[82],"computation":[83],"across":[84],"GPUs":[85],"ensures":[87],"full":[88],"retention,":[90],"maximizing":[91],"regardless":[93],"skewness.":[96],"this":[99],"strategic":[102],"row-":[103],"column-wise":[105],"decomposition":[106],"matrices.":[109],"This":[110],"reduces":[111],"idle":[112],"time":[113,150],"avoids":[115],"bottlenecks":[116],"caused":[117],"by":[118,127],"assignments.":[121],"Furthermore,":[122],"minimizes":[124],"kernel":[125],"launches":[126],"fusing":[128],"decomposed":[129],"computations,":[131],"further":[132],"improving":[133],"throughput.":[134],"evaluate":[136],"against":[138],"DeepSpeed":[139],"demonstrating":[143],"speedups":[144],"up":[146],"6.4\u00d7":[148],"first":[152],"(TTFT).":[154],"Our":[155],"show":[157],"when":[159],"properly":[160],"applied":[161],"experts,":[163],"is":[166],"viable":[168],"effective":[170],"strategy":[171],"efficient":[173],"inference.":[175]},"counts_by_year":[{"year":2026,"cited_by_count":2}],"updated_date":"2026-05-06T08:25:59.206177","created_date":"2025-10-10T00:00:00"}