{"id":"https://openalex.org/W7225000280","doi":"https://doi.org/10.5281/zenodo.23840457","title":"TorchTPU \u00b7 Gemma 2B greedy decoding on TPU v5e-1: Boosting inference to 21,645 tok/s","display_name":"TorchTPU \u00b7 Gemma 2B greedy decoding on TPU v5e-1: Boosting inference to 21,645 tok/s","publication_year":2026,"publication_date":"2026-10-31","ids":{"openalex":"https://openalex.org/W7225000280","doi":"https://doi.org/10.5281/zenodo.23840457"},"language":"en","primary_location":{"id":"doi:10.5281/zenodo.23840457","doi":"https://doi.org/10.5281/zenodo.23840457","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23840457","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"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"},"volume":null,"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.5281/zenodo.23840457","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5036401688","display_name":"Rubens de Almeida Zimbres","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"de Almeida Zimbres, Rubens","raw_affiliation_strings":["Intellimetri"],"raw_orcid":"https://orcid.org/0000-0003-4765-2885","affiliations":[{"raw_affiliation_string":"Intellimetri","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5036401688"],"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.9398999810218811,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.9398999810218811,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10181","display_name":"Natural Language Processing Techniques","score":0.012600000016391277,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.011699999682605267,"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/greedy-decoding","display_name":"greedy decoding","score":0.9240000247955322},{"id":"https://openalex.org/keywords/token-throughput","display_name":"token throughput","score":0.5289999842643738},{"id":"https://openalex.org/keywords/key-value-cache","display_name":"key-value cache","score":0.6890000104904175},{"id":"https://openalex.org/keywords/time-to-first-token","display_name":"time-to-first-token","score":0.6779999732971191}],"concepts":[{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.8407999873161316},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7520999908447266},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.5839999914169312},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.5254999995231628},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5034000277519226},{"id":"https://openalex.org/C169590947","wikidata":"https://www.wikidata.org/wiki/Q47506","display_name":"Compiler","level":2,"score":0.4530999958515167},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.4092999994754791},{"id":"https://openalex.org/C46686674","wikidata":"https://www.wikidata.org/wiki/Q466303","display_name":"Boosting (machine learning)","level":2,"score":0.391400009393692},{"id":"https://openalex.org/C2778476105","wikidata":"https://www.wikidata.org/wiki/Q628539","display_name":"Workload","level":2,"score":0.35109999775886536},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.3352000117301941},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.3228999972343445},{"id":"https://openalex.org/C165005293","wikidata":"https://www.wikidata.org/wiki/Q1074500","display_name":"Chip","level":2,"score":0.31470000743865967},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.30070000886917114},{"id":"https://openalex.org/C103088060","wikidata":"https://www.wikidata.org/wiki/Q1062839","display_name":"Error detection and correction","level":2,"score":0.2644999921321869},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.2635999917984009}],"mesh":[],"locations_count":2,"locations":[{"id":"doi:10.5281/zenodo.23840457","doi":"https://doi.org/10.5281/zenodo.23840457","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23840457","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"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"},{"id":"doi:10.5281/zenodo.23840458","doi":"https://doi.org/10.5281/zenodo.23840458","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23840458","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"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"}],"best_oa_location":{"id":"doi:10.5281/zenodo.23840457","doi":"https://doi.org/10.5281/zenodo.23840457","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23840457","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"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":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7167483051","https://openalex.org/W7169777772","https://openalex.org/W7167456272","https://openalex.org/W7162454159","https://openalex.org/W7169578937","https://openalex.org/W7223028986","https://openalex.org/W7203570064","https://openalex.org/W7172224523","https://openalex.org/W7154961471","https://openalex.org/W7167900054"],"abstract_inverted_index":{"TorchTPU":[0],"is":[1],"a":[2,13,52,65,90,103,121,157],"PyTorch":[3,7],"backend":[4],"that":[5],"runs":[6],"programs":[8],"on":[9,25,72,234],"Google":[10],"TPUs":[11],"through":[12],"nativetpu":[14],"device.":[15],"This":[16],"report":[17,196,217],"measures":[18],"greedy":[19],"text":[20],"generation":[21],"with":[22,95,138,156],"Gemma":[23],"2B":[24],"one":[26,87],"TPU":[27],"v5e":[28],"chip":[29,75],"usingthe":[30],"public":[31],"release":[32],"torch_tpu":[33],"0.2.0.":[34],"No":[35],"serving":[36],"engine":[37],"exists":[38],"for":[39,60,229],"TorchTPU,":[40],"so":[41],"the":[42,44,48,70,73,92,116,139,146,149,186,195,201,204,211,226],"model,":[43],"keyand":[45],"value":[46],"cache,":[47],"decoding":[49],"loop,":[50],"and":[51,78,130,172,192,225],"decode":[53,93,133],"attention":[54],"kernel":[55,118],"in":[56,190],"Pallas":[57,117],"were":[58],"written":[59],"thisstudy.":[61],"The":[62,178,216],"workload":[63],"repeats":[64],"previous":[66],"vLLM":[67],"experiment":[68],"by":[69,102],"author":[71],"same":[74],"class:512":[76],"sequences":[77],"128":[79,101],"generated":[80,128],"tokens":[81,110,153,174,183],"per":[82,111,154,184],"sequence.":[83],"Four":[84],"experiments":[85],"add":[86],"change":[88],"at":[89,99,207],"time.Compiling":[91],"step":[94],"torch.compile":[96],"raises":[97],"throughput":[98],"batch":[100,114,158],"factor":[104],"of":[105,163,169,200,210],"5.1,from":[106],"2,095":[107],"to":[108,160],"10,624":[109],"second.":[112],"At":[113],"512,":[115],"registered":[119],"as":[120],"PyTorchoperation":[122],"adds":[123,142],"5.6":[124],"percent":[125,209],"over":[126],"compiler":[127],"attention,":[129],"running":[131],"eight":[132],"steps":[134],"insideone":[135],"compiled":[136],"program":[137],"scan":[140],"operator":[141],"21.1":[143],"percent.":[144],"With":[145],"real":[147],"weights":[148],"finalconfiguration":[150],"reaches":[151],"21,645.8":[152],"second,":[155],"time":[159],"first":[161],"token":[162,167],"222":[164],"milliseconds,an":[165],"inter":[166],"latency":[168],"22.1":[170],"milliseconds,":[171],"identical":[173],"across":[175],"repeated":[176],"runs.":[177],"vLLMexperiment":[179],"had":[180],"reached":[181],"19,511":[182],"second;":[185],"two":[187],"systems":[188],"differ":[189],"scheduling":[191],"cachedesign,":[193],"which":[194],"lists.":[197],"An":[198],"estimate":[199],"arithmetic":[202],"places":[203],"final":[205],"configuration":[206],"about55":[208],"chip\u2019s":[212],"peak":[213],"bfloat16":[214],"compute.":[215],"explains":[218],"each":[219,222],"component,":[220],"how":[221],"resultwas":[223],"verified,":[224],"practical":[227],"lessons":[228],"engineers":[230],"who":[231],"run":[232],"inference":[233],"TorchTPU.":[235]},"counts_by_year":[],"updated_date":"2026-11-02T09:45:03.842371","created_date":"2026-11-01T00:00:00"}