{"id":"https://openalex.org/W7168365243","doi":"https://doi.org/10.48550/arxiv.2607.12505","title":"Realizable N:M Sparse Transformer Inference via Search-Kernel Co-Design","display_name":"Realizable N:M Sparse Transformer Inference via Search-Kernel Co-Design","publication_year":2026,"publication_date":"2026-07-14","ids":{"openalex":"https://openalex.org/W7168365243","doi":"https://doi.org/10.48550/arxiv.2607.12505"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2607.12505","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.12505","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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.2607.12505","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5140960541","display_name":"Yiming Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Yiming","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140752271","display_name":"Wenqi Lou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lou, Wenqi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140757993","display_name":"Zhiguang Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Zhiguang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140812507","display_name":"Zhiwei Ke","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ke, Zhiwei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140919355","display_name":"Fengrui Zuo","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zuo, Fengrui","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140762949","display_name":"Chao Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Chao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5140929412","display_name":"Xuehai Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Xuehai","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/T10036","display_name":"Advanced Neural Network Applications","score":0.25589999556541443,"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":0.25589999556541443,"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/T11992","display_name":"CCD and CMOS Imaging Sensors","score":0.24770000576972961,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10531","display_name":"Advanced Vision and Imaging","score":0.08479999750852585,"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/dataflow","display_name":"Dataflow","score":0.6549999713897705},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.6273999810218811},{"id":"https://openalex.org/keywords/speedup","display_name":"Speedup","score":0.5899999737739563},{"id":"https://openalex.org/keywords/latency","display_name":"Latency (audio)","score":0.4756999909877777},{"id":"https://openalex.org/keywords/scheduling","display_name":"Scheduling (production processes)","score":0.4575999975204468},{"id":"https://openalex.org/keywords/cuda","display_name":"CUDA","score":0.45489999651908875},{"id":"https://openalex.org/keywords/sparse-matrix","display_name":"Sparse matrix","score":0.45419999957084656},{"id":"https://openalex.org/keywords/software","display_name":"Software","score":0.41659998893737793},{"id":"https://openalex.org/keywords/kernel","display_name":"Kernel (algebra)","score":0.39309999346733093}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8361999988555908},{"id":"https://openalex.org/C96324660","wikidata":"https://www.wikidata.org/wiki/Q205446","display_name":"Dataflow","level":2,"score":0.6549999713897705},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.6509000062942505},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6273999810218811},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.5899999737739563},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.4756999909877777},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.4575999975204468},{"id":"https://openalex.org/C2778119891","wikidata":"https://www.wikidata.org/wiki/Q477690","display_name":"CUDA","level":2,"score":0.45489999651908875},{"id":"https://openalex.org/C56372850","wikidata":"https://www.wikidata.org/wiki/Q1050404","display_name":"Sparse matrix","level":3,"score":0.45419999957084656},{"id":"https://openalex.org/C2777904410","wikidata":"https://www.wikidata.org/wiki/Q7397","display_name":"Software","level":2,"score":0.41659998893737793},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.3995000123977661},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.39309999346733093},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.38100001215934753},{"id":"https://openalex.org/C2780513914","wikidata":"https://www.wikidata.org/wiki/Q18210350","display_name":"Bottleneck","level":2,"score":0.36970001459121704},{"id":"https://openalex.org/C78766204","wikidata":"https://www.wikidata.org/wiki/Q555032","display_name":"Multi-core processor","level":2,"score":0.3603000044822693},{"id":"https://openalex.org/C169590947","wikidata":"https://www.wikidata.org/wiki/Q47506","display_name":"Compiler","level":2,"score":0.33980000019073486},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.3089999854564667},{"id":"https://openalex.org/C42935608","wikidata":"https://www.wikidata.org/wiki/Q190411","display_name":"Field-programmable gate array","level":2,"score":0.290800005197525},{"id":"https://openalex.org/C58013763","wikidata":"https://www.wikidata.org/wiki/Q5754574","display_name":"High-level synthesis","level":3,"score":0.2856999933719635},{"id":"https://openalex.org/C189930140","wikidata":"https://www.wikidata.org/wiki/Q1112878","display_name":"CAS latency","level":4,"score":0.28220000863075256},{"id":"https://openalex.org/C76752949","wikidata":"https://www.wikidata.org/wiki/Q7607499","display_name":"Stencil","level":2,"score":0.28209999203681946},{"id":"https://openalex.org/C79581498","wikidata":"https://www.wikidata.org/wiki/Q1367530","display_name":"Suite","level":2,"score":0.2802000045776367},{"id":"https://openalex.org/C43126263","wikidata":"https://www.wikidata.org/wiki/Q128751","display_name":"Source code","level":2,"score":0.2754000127315521},{"id":"https://openalex.org/C173801870","wikidata":"https://www.wikidata.org/wiki/Q201413","display_name":"Heuristic","level":2,"score":0.2732999920845032},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.2623000144958496},{"id":"https://openalex.org/C200833197","wikidata":"https://www.wikidata.org/wiki/Q333707","display_name":"Compile time","level":3,"score":0.2565000057220459},{"id":"https://openalex.org/C177774035","wikidata":"https://www.wikidata.org/wiki/Q1246948","display_name":"Granularity","level":2,"score":0.2535000145435333}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2607.12505","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.12505","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2607.12505","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.12505","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"display_name":"Industry, innovation and infrastructure","id":"https://metadata.un.org/sdg/9","score":0.40799787640571594}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Vision":[0],"Transformers":[1],"(ViTs)":[2],"achieve":[3],"strong":[4],"accuracy":[5,161,165],"but":[6,45],"incur":[7],"high":[8],"inference":[9],"latency.":[10],"Semi-structured":[11],"N:M":[12,78,90],"sparsity":[13,66,117,139],"can":[14],"reduce":[15],"arithmetic":[16,43],"cost,":[17],"yet":[18],"its":[19],"theoretical":[20],"savings":[21],"often":[22],"fail":[23],"to":[24,107,132],"translate":[25],"into":[26,98],"proportional":[27],"end-to-end":[28,121],"speedups":[29],"on":[30,42,47,142],"modern":[31],"GPUs.":[32],"This":[33],"mismatch":[34],"arises":[35],"because":[36],"deployment":[37],"latency":[38,122,156,169],"depends":[39],"not":[40],"only":[41],"reduction":[44],"also":[46],"execution":[48,64],"regularity":[49],"and":[50,65,102,136,146,162],"hardware":[51,84],"scheduling":[52],"under":[53,119,166],"sparsity.":[54],"Achieving":[55],"practical":[56],"acceleration,":[57],"therefore,":[58],"requires":[59],"coordinated":[60],"design":[61,87],"across":[62],"sparse":[63,79,91,96],"configuration.":[67],"To":[68],"this":[69],"end,":[70],"we":[71,86,114],"propose":[72],"a":[73,125],"hardware-software":[74],"co-design":[75],"framework":[76,152],"for":[77],"ViT":[80],"inference.":[81],"On":[82,110],"the":[83,111,151,167],"side,":[85,113],"MD-SpMM,":[88],"an":[89],"CUDA":[92],"kernel":[93],"that":[94,150],"reorganizes":[95],"GEMM":[97],"micro-dense,":[99],"Tensor-Core-aligned":[100],"dataflow":[101],"uses":[103],"inference-aware":[104],"adaptive":[105],"parallelism":[106],"sustain":[108],"utilization.":[109],"software":[112],"perform":[115],"layer-wise":[116],"search":[118,128],"explicit":[120],"budgets":[123],"using":[124],"three-stage":[126],"heuristic":[127],"with":[129],"constraint":[130],"relaxation":[131],"avoid":[133],"premature":[134],"convergence":[135],"enable":[137],"deployment-aware":[138],"allocation.":[140],"Experiments":[141],"multiple":[143],"ViT/Swin":[144],"models":[145],"GPU":[147],"platforms":[148],"show":[149],"achieves":[153],"over":[154],"2.2x":[155],"speedup":[157],"while":[158],"maintaining":[159],"comparable":[160],"delivering":[163],"superior":[164],"same":[168],"constraint.":[170],"The":[171],"source":[172],"code":[173],"is":[174],"publicly":[175],"available":[176],"at":[177],"https://github.com/liuganhuo/realizable-nm-sparse-transformer.":[178]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-07-16T00:00:00"}