{"id":"https://openalex.org/W7164826052","doi":"https://doi.org/10.48550/arxiv.2606.13740","title":"Efficient On-Device Diffusion LLM Inference with Mobile NPU","display_name":"Efficient On-Device Diffusion LLM Inference with Mobile NPU","publication_year":2026,"publication_date":"2026-06-11","ids":{"openalex":"https://openalex.org/W7164826052","doi":"https://doi.org/10.48550/arxiv.2606.13740"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.13740","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.13740","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.2606.13740","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5067496010","display_name":"Tuowei Wang","orcid":"https://orcid.org/0009-0006-8272-8151"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Tuowei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5069456711","display_name":"\u5b59\u5ef6\u51e1","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Yanfan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138645189","display_name":"Ju Ren","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ren, Ju","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":0.39719998836517334,"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":0.39719998836517334,"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":0.060499999672174454,"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/T11948","display_name":"Machine Learning in Materials Science","score":0.052299998700618744,"subfield":{"id":"https://openalex.org/subfields/2505","display_name":"Materials Chemistry"},"field":{"id":"https://openalex.org/fields/25","display_name":"Materials Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.8177000284194946},{"id":"https://openalex.org/keywords/cache","display_name":"Cache","score":0.6068999767303467},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5845999717712402},{"id":"https://openalex.org/keywords/latency","display_name":"Latency (audio)","score":0.5748999714851379},{"id":"https://openalex.org/keywords/decoding-methods","display_name":"Decoding methods","score":0.49149999022483826},{"id":"https://openalex.org/keywords/computation","display_name":"Computation","score":0.48590001463890076},{"id":"https://openalex.org/keywords/mobile-device","display_name":"Mobile device","score":0.4636000096797943},{"id":"https://openalex.org/keywords/dram","display_name":"Dram","score":0.3878999948501587},{"id":"https://openalex.org/keywords/transfer","display_name":"Transfer (computing)","score":0.3822000026702881}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8529000282287598},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.8177000284194946},{"id":"https://openalex.org/C115537543","wikidata":"https://www.wikidata.org/wiki/Q165596","display_name":"Cache","level":2,"score":0.6068999767303467},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5845999717712402},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.5748999714851379},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.534500002861023},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.49149999022483826},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.48590001463890076},{"id":"https://openalex.org/C186967261","wikidata":"https://www.wikidata.org/wiki/Q5082128","display_name":"Mobile device","level":2,"score":0.4636000096797943},{"id":"https://openalex.org/C7366592","wikidata":"https://www.wikidata.org/wiki/Q1255620","display_name":"Dram","level":2,"score":0.3878999948501587},{"id":"https://openalex.org/C2776175482","wikidata":"https://www.wikidata.org/wiki/Q1195816","display_name":"Transfer (computing)","level":2,"score":0.3822000026702881},{"id":"https://openalex.org/C189783530","wikidata":"https://www.wikidata.org/wiki/Q352090","display_name":"CPU cache","level":3,"score":0.374099999666214},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.3492000102996826},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.32910001277923584},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.3212999999523163},{"id":"https://openalex.org/C144543869","wikidata":"https://www.wikidata.org/wiki/Q2738570","display_name":"Mobile computing","level":2,"score":0.31279999017715454},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.30809998512268066},{"id":"https://openalex.org/C2778476105","wikidata":"https://www.wikidata.org/wiki/Q628539","display_name":"Workload","level":2,"score":0.30300000309944153},{"id":"https://openalex.org/C141603448","wikidata":"https://www.wikidata.org/wiki/Q134830","display_name":"Prefix","level":2,"score":0.29100000858306885},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.2879999876022339},{"id":"https://openalex.org/C32638748","wikidata":"https://www.wikidata.org/wiki/Q5437051","display_name":"Fast path","level":4,"score":0.287200003862381},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2840999960899353},{"id":"https://openalex.org/C147168706","wikidata":"https://www.wikidata.org/wiki/Q1457734","display_name":"Recurrent neural network","level":3,"score":0.28299999237060547},{"id":"https://openalex.org/C38556500","wikidata":"https://www.wikidata.org/wiki/Q13404475","display_name":"Cache algorithms","level":4,"score":0.28279998898506165},{"id":"https://openalex.org/C163294075","wikidata":"https://www.wikidata.org/wiki/Q581861","display_name":"Noise reduction","level":2,"score":0.2752000093460083},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.2732999920845032},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.2703000009059906},{"id":"https://openalex.org/C78766204","wikidata":"https://www.wikidata.org/wiki/Q555032","display_name":"Multi-core processor","level":2,"score":0.2655999958515167},{"id":"https://openalex.org/C2779808786","wikidata":"https://www.wikidata.org/wiki/Q6664603","display_name":"Locality","level":2,"score":0.2651999890804291},{"id":"https://openalex.org/C46637626","wikidata":"https://www.wikidata.org/wiki/Q6693015","display_name":"Low latency (capital markets)","level":2,"score":0.26350000500679016},{"id":"https://openalex.org/C176649486","wikidata":"https://www.wikidata.org/wiki/Q2308807","display_name":"Memory management","level":3,"score":0.25690001249313354},{"id":"https://openalex.org/C138236772","wikidata":"https://www.wikidata.org/wiki/Q25098575","display_name":"Edge device","level":3,"score":0.2554999887943268},{"id":"https://openalex.org/C2777402240","wikidata":"https://www.wikidata.org/wiki/Q6783436","display_name":"Masking (illustration)","level":2,"score":0.2547999918460846}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.13740","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.13740","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.2606.13740","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.13740","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":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Diffusion":[0],"large":[1],"language":[2],"models":[3],"(dLLMs)":[4],"accelerate":[5],"generation":[6,179,194],"by":[7,181],"denoising":[8,22],"multiple":[9],"tokens":[10,121,128],"in":[11,107],"parallel,":[12],"making":[13],"them":[14,41],"attractive":[15],"for":[16,79],"latency-sensitive":[17],"mobile":[18,94],"inference.":[19],"However,":[20],"repeated":[21],"introduces":[23],"substantial":[24],"computation":[25,152],"on":[26,82],"smartphones.":[27,83],"Mobile":[28],"neural":[29],"processing":[30],"units":[31],"(NPUs)":[32],"offer":[33],"high-throughput":[34],"dense":[35,135],"matrix":[36],"computation,":[37],"but":[38],"efficiently":[39],"exploiting":[40],"remains":[42],"challenging:":[43],"token":[44,50],"commitment":[45],"shrinks":[46],"per-block":[47],"effective":[48],"workloads,":[49],"revision":[51],"complicates":[52],"KV":[53,189],"cache":[54,190],"reuse,":[55,191],"and":[56,64,125,146,156,166,173],"limited":[57],"NPU-visible":[58,143],"address":[59,144],"space":[60],"incurs":[61],"costly":[62],"remapping":[63,155],"data":[65,148],"transfer":[66,157],"overheads.":[67,158],"In":[68],"this":[69],"paper,":[70],"we":[71],"propose":[72],"llada.cpp,":[73],"the":[74,90,104,184],"first":[75],"NPU-aware":[76],"inference":[77,88],"framework":[78,165],"accelerating":[80],"dLLMs":[81],"llada.cpp":[84,161,176],"aligns":[85],"block-wise":[86],"dLLM":[87,174],"with":[89,111,150,187],"execution":[91],"characteristics":[92],"of":[93],"NPUs":[95],"through":[96,129],"three":[97],"techniques.":[98],"(1)":[99],"Multi-Block":[100],"Speculative":[101],"Decoding":[102],"fills":[103],"shrinking":[105],"workload":[106],"late-stage":[108],"current-block":[109],"decoding":[110],"speculative":[112],"future-block":[113],"tokens.":[114],"(2)":[115],"Dual-Path":[116],"Progressive":[117],"Revision":[118],"keeps":[119],"committed":[120],"revisable":[122],"until":[123],"stable":[124],"refreshes":[126],"unstable":[127],"a":[130],"CPU-side":[131],"path":[132],"without":[133],"stalling":[134],"NPU":[136,151],"execution.":[137],"(3)":[138],"Swap-Optimized":[139],"Memory":[140],"Runtime":[141],"compacts":[142],"layouts":[145],"overlaps":[147],"staging":[149],"to":[153],"reduce":[154],"We":[159],"implement":[160],"as":[162],"an":[163],"end-to-end":[164],"evaluate":[167],"it":[168],"across":[169],"diverse":[170],"hardware":[171],"platforms":[172],"workloads.":[175],"reduces":[177],"LLaDA-8B":[178],"latency":[180],"17x-42x":[182],"over":[183],"CPU":[185],"baseline":[186],"prefix":[188],"while":[192],"preserving":[193],"quality.":[195]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-06-16T00:00:00"}