{"id":"https://openalex.org/W4411950023","doi":"https://doi.org/10.1109/tc.2025.3585339","title":"<i>FlashDecoding++Next</i>: High Throughput LLM Inference With Latency and Memory Optimization","display_name":"<i>FlashDecoding++Next</i>: High Throughput LLM Inference With Latency and Memory Optimization","publication_year":2025,"publication_date":"2025-07-02","ids":{"openalex":"https://openalex.org/W4411950023","doi":"https://doi.org/10.1109/tc.2025.3585339"},"language":"en","primary_location":{"id":"doi:10.1109/tc.2025.3585339","is_oa":false,"landing_page_url":"https://doi.org/10.1109/tc.2025.3585339","pdf_url":null,"source":{"id":"https://openalex.org/S157670870","display_name":"IEEE Transactions on Computers","issn_l":"0018-9340","issn":["0018-9340","0016-9340","1557-9956","2326-3814"],"is_oa":false,"is_in_doaj":false,"is_core":true,"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"journal"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Transactions on Computers","raw_type":"journal-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":null,"display_name":"Guohao Dai","orcid":"https://orcid.org/0000-0002-8464-0130"},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Guohao Dai","raw_affiliation_strings":["School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China","Qing Yuan Research Institute, Shanghai Jiao Tong University, China"],"raw_orcid":"https://orcid.org/0000-0002-8464-0130","affiliations":[{"raw_affiliation_string":"School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China","institution_ids":["https://openalex.org/I183067930"]},{"raw_affiliation_string":"Qing Yuan Research Institute, Shanghai Jiao Tong University, China","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5014087671","display_name":"Ke Hong","orcid":"https://orcid.org/0000-0002-5768-6037"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Ke Hong","raw_affiliation_strings":["Department of Electronic Engineering, Tsinghua University, Beijing, China","Department of Electronic Engineering, Tsinghua University, China"],"raw_orcid":"https://orcid.org/0000-0002-5768-6037","affiliations":[{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]},{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5076000240","display_name":"Qiuli Mao","orcid":"https://orcid.org/0009-0004-8777-2579"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Qiuli Mao","raw_affiliation_strings":["Infinigence-AI, Shanghai, China","Department of Electronic Engineering, Tsinghua University, China"],"raw_orcid":"https://orcid.org/0009-0004-8777-2579","affiliations":[{"raw_affiliation_string":"Infinigence-AI, Shanghai, China","institution_ids":[]},{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5114246744","display_name":"Xiuhong Li","orcid":"https://orcid.org/0000-0002-4896-121X"},"institutions":[{"id":"https://openalex.org/I20231570","display_name":"Peking University","ror":"https://ror.org/02v51f717","country_code":"CN","type":"education","lineage":["https://openalex.org/I20231570"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Xiuhong Li","raw_affiliation_strings":["Infinigence-AI, Shanghai, China","Peking University, China"],"raw_orcid":"https://orcid.org/0000-0002-4896-121X","affiliations":[{"raw_affiliation_string":"Infinigence-AI, Shanghai, China","institution_ids":[]},{"raw_affiliation_string":"Peking University, China","institution_ids":["https://openalex.org/I20231570"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Jiaming Xu","orcid":"https://orcid.org/0009-0000-7000-6537"},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Jiaming Xu","raw_affiliation_strings":["Qing Yuan Research Institute, Shanghai Jiao Tong University, Shanghai, China","Infinigence-AI"],"raw_orcid":"https://orcid.org/0009-0000-7000-6537","affiliations":[{"raw_affiliation_string":"Qing Yuan Research Institute, Shanghai Jiao Tong University, Shanghai, China","institution_ids":["https://openalex.org/I183067930"]},{"raw_affiliation_string":"Infinigence-AI","institution_ids":[]}]},{"author_position":"middle","author":{"id":null,"display_name":"Haofeng Huang","orcid":"https://orcid.org/0009-0002-7253-1364"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Haofeng Huang","raw_affiliation_strings":["Department of Electronic Engineering, Tsinghua University, Beijing, China","Department of Electronic Engineering, Tsinghua University, China"],"raw_orcid":"https://orcid.org/0009-0002-7253-1364","affiliations":[{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]},{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Hongtu Xia","orcid":"https://orcid.org/0009-0009-3340-2210"},"institutions":[{"id":"https://openalex.org/I20231570","display_name":"Peking University","ror":"https://ror.org/02v51f717","country_code":"CN","type":"education","lineage":["https://openalex.org/I20231570"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Hongtu Xia","raw_affiliation_strings":["Peking University, Beijing, China","Infinigence-AI"],"raw_orcid":"https://orcid.org/0009-0009-3340-2210","affiliations":[{"raw_affiliation_string":"Peking University, Beijing, China","institution_ids":["https://openalex.org/I20231570"]},{"raw_affiliation_string":"Infinigence-AI","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5086217226","display_name":"Xuefei Ning","orcid":"https://orcid.org/0000-0003-2209-8312"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Xuefei Ning","raw_affiliation_strings":["Department of Electronic Engineering, Tsinghua University, Beijing, China","Department of Electronic Engineering, Tsinghua University, China"],"raw_orcid":"https://orcid.org/0000-0003-2209-8312","affiliations":[{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]},{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5049910854","display_name":"Shengen Yan","orcid":"https://orcid.org/0009-0005-3858-7972"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Shengen Yan","raw_affiliation_strings":["Department of Electronic Engineering, Tsinghua University, Beijing, China","Department of Electronic Engineering, Tsinghua University, China"],"raw_orcid":"https://orcid.org/0009-0005-3858-7972","affiliations":[{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]},{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100604860","display_name":"Yun Liang","orcid":"https://orcid.org/0000-0002-9076-7998"},"institutions":[{"id":"https://openalex.org/I20231570","display_name":"Peking University","ror":"https://ror.org/02v51f717","country_code":"CN","type":"education","lineage":["https://openalex.org/I20231570"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yun Liang","raw_affiliation_strings":["Peking University, Beijing, China","Peking University, China"],"raw_orcid":"https://orcid.org/0000-0002-9076-7998","affiliations":[{"raw_affiliation_string":"Peking University, Beijing, China","institution_ids":["https://openalex.org/I20231570"]},{"raw_affiliation_string":"Peking University, China","institution_ids":["https://openalex.org/I20231570"]}]},{"author_position":"last","author":{"id":null,"display_name":"Yu Wang","orcid":"https://orcid.org/0000-0001-6108-5157"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yu Wang","raw_affiliation_strings":["Department of Electronic Engineering, Tsinghua University, Beijing, China","Department of Electronic Engineering, Tsinghua University, China"],"raw_orcid":"https://orcid.org/0000-0001-6108-5157","affiliations":[{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]},{"raw_affiliation_string":"Department of Electronic Engineering, Tsinghua University, China","institution_ids":["https://openalex.org/I99065089"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.9389,"has_fulltext":false,"cited_by_count":1,"citation_normalized_percentile":{"value":0.7715501,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":94,"max":97},"biblio":{"volume":"74","issue":"10","first_page":"3263","last_page":"3276"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11181","display_name":"Advanced Data Storage Technologies","score":0.9739999771118164,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T11181","display_name":"Advanced Data Storage Technologies","score":0.9739999771118164,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/computer-science","display_name":"Computer science","score":0.6998533010482788},{"id":"https://openalex.org/keywords/latency","display_name":"Latency (audio)","score":0.5831940770149231},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5780060291290283},{"id":"https://openalex.org/keywords/throughput","display_name":"Throughput","score":0.5559014678001404},{"id":"https://openalex.org/keywords/parallel-computing","display_name":"Parallel computing","score":0.5193415880203247},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.19470137357711792},{"id":"https://openalex.org/keywords/operating-system","display_name":"Operating system","score":0.15331873297691345},{"id":"https://openalex.org/keywords/telecommunications","display_name":"Telecommunications","score":0.08985403180122375}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6998533010482788},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.5831940770149231},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5780060291290283},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.5559014678001404},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.5193415880203247},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.19470137357711792},{"id":"https://openalex.org/C111919701","wikidata":"https://www.wikidata.org/wiki/Q9135","display_name":"Operating system","level":1,"score":0.15331873297691345},{"id":"https://openalex.org/C76155785","wikidata":"https://www.wikidata.org/wiki/Q418","display_name":"Telecommunications","level":1,"score":0.08985403180122375},{"id":"https://openalex.org/C555944384","wikidata":"https://www.wikidata.org/wiki/Q249","display_name":"Wireless","level":2,"score":0.0}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/tc.2025.3585339","is_oa":false,"landing_page_url":"https://doi.org/10.1109/tc.2025.3585339","pdf_url":null,"source":{"id":"https://openalex.org/S157670870","display_name":"IEEE Transactions on Computers","issn_l":"0018-9340","issn":["0018-9340","0016-9340","1557-9956","2326-3814"],"is_oa":false,"is_in_doaj":false,"is_core":true,"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"journal"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Transactions on Computers","raw_type":"journal-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":9,"referenced_works":["https://openalex.org/W2946609015","https://openalex.org/W2964110616","https://openalex.org/W2996908057","https://openalex.org/W4321636575","https://openalex.org/W4384561707","https://openalex.org/W4385245566","https://openalex.org/W4387321091","https://openalex.org/W4387500346","https://openalex.org/W4394595621"],"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/W3128807919","https://openalex.org/W3176411177"],"abstract_inverted_index":{"As":[0],"the":[1,12,26,30,33,66,149,189,231,244,249,260,270,298],"Large":[2],"Language":[3],"Model":[4],"(LLM)":[5],"becomes":[6],"increasingly":[7],"important":[8],"in":[9,38,69,83,101,194,225],"various":[10],"domains,":[11],"performance":[13,96],"of":[14,74,79,116,124,262,265],"LLM":[15,21,41,84,195],"inference":[16,85,119,139,250],"is":[17,181],"crucial":[18],"to":[19,62,87,91,121,172,184,227,252,267,278,282,292,297,312],"massive":[20],"applications.":[22],"However,":[23],"centering":[24],"around":[25],"computational":[27],"efficiency":[28],"and":[29,94,144,186,191,238,307,315],"memory":[31,130,240,274],"utilization,":[32],"following":[34],"challenges":[35],"remain":[36],"unsolved":[37],"achieving":[39],"high-throughput":[40,138],"inference:":[42],"(1)":[43,155],"Synchronous":[44],"partial":[45,58,169],"softmax":[46,49,59,157,170],"update.":[47],"The":[48,77,273],"operation":[50,55],"requires":[51],"a":[52,137,163,178],"synchronous":[53],"update":[54],"among":[56],"each":[57],"result,":[60],"leading":[61,90,183],"~20%":[63],"overheads":[64],"for":[65,167,188,230],"attention":[67],"computation":[68,73,93],"LLMs.":[70,320],"(2)":[71,198],"Under-utilized":[72],"flat":[75,209,232],"GEMM.":[76],"shape":[78],"matrices":[80],"performing":[81],"GEMM":[82,200,233],"tends":[86],"be":[88,283],"flat,":[89],"under-utilized":[92],"50%":[95],"loss":[97],"after":[98],"padding":[99],"zeros":[100],"previous":[102],"designs":[103],"(e.g.,":[104],"cuBLAS,":[105],"CUTLASS,":[106],"etc.).":[107],"(3)":[108,235],"Memory":[109],"redundancy":[110],"caused":[111],"by":[112],"activations.":[113],"Dynamic":[114],"allocation":[115],"activations":[117],"during":[118],"leads":[120],"redundant":[122],"storage":[123,266],"useless":[125],"variables,":[126],"bringing":[127],"22%":[128],"more":[129],"consumption.":[131],"<p":[132],"xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"":[133],"xmlns:xlink=\"http://www.w3.org/1999/xlink\">We":[134],"present":[135],"<i>FlashDecoding</i>++<i>Next</i>,":[136],"engine":[140],"supporting":[141],"mainstream":[142,319],"LLMs":[143],"hardware":[145],"backends.":[146],"To":[147],"tackle":[148],"above":[150],"challenges,":[151],"<i>FlashDecoding</i>++<i>Next</i>":[152,161,205,242,285,304],"creatively":[153],"proposes:":[154],"Asynchronous":[156],"with":[158,202,211],"unified":[159,164,239],"maximum.":[160],"introduces":[162],"maximum":[165],"technique":[166],"different":[168,212,263],"computations":[171],"avoid":[173],"synchronization.":[174],"Based":[175,255],"on":[176,256,318],"this,":[177],"fine-grained":[179],"pipelining":[180],"proposed,":[182],"1.18\u00d7":[185],"1.14\u00d7":[187],"<i>prefill</i>":[190],"<i>decode</i>":[192],"phases":[193],"inference,":[196],"respectively.":[197],"Flat":[199],"optimization":[201,275],"double":[203,220],"buffering.":[204],"points":[206],"out":[207],"that":[208],"GEMMs":[210],"shapes":[213],"face":[214],"varied":[215],"bottlenecks.":[216],"Then,":[217],"techniques":[218],"like":[219],"buffering":[221],"are":[222],"introduced,":[223],"resulting":[224],"up":[226,277,291],"52%":[228],"speedup":[229],"operation.":[234],"Buffer":[236],"reusing":[237,271],"management.":[241],"reuses":[243],"pre-allocated":[245],"activation":[246],"buffers":[247],"throughout":[248],"process":[251],"remove":[253],"redundancy.":[254],"that,":[257],"we":[258],"unify":[259],"management":[261],"types":[264],"further":[268],"exploit":[269],"opportunity.":[272],"enables":[276],"1.57\u00d7":[279],"longer":[280],"sequence":[281],"processed.":[284],"demonstrates":[286],"remarkable":[287],"throughput":[288,295,310],"improvement,":[289],"delivering":[290],"68.88\u00d7":[293],"higher":[294,309],"compared":[296,311],"HuggingFace":[299],"[1]":[300],"implementation.":[301],"On":[302],"average,":[303],"achieves":[305],"1.25\u00d7":[306],"1.46\u00d7":[308],"vLLM":[313],"[2]":[314],"TensorRT-LLM":[316],"[3]":[317]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-03-27T05:58:40.876381","created_date":"2025-07-03T00:00:00"}