Chii chaitika
Vatsvagiri vakaunza GLANCE, imwe-pass block drafter yekufungidzira decoding mumamodeli emutauro wechiratidzo. Iyo nzira inoverenga iyo inonangwa modhi yakatosanganiswa yekuona uye yemutauro inomiririra, inogadzira bhuroka rema tokens mune imwe yekupfuura pass, uye ine iyo yakanangwa modhi inosimbisa avo vanokwikwidza mukupasa kumwe chete. Vanyori vanoshuma kuti zvakaongororwa zvinokurudzira zvakaburitsa kukara decoding chaizvo uye kuti GLANCE yakasvika kusvika 2.93 × kumhanya kweautoregressive decoding pasi pemamiriro avo ebvunzo akataurwa.
Bepa rakatumirwa kuarXiv musi waAug. 31 rinounza GLANCE, iyo yainotsanangura seimwe-pass block drafter yekufungidzira decoding mumamodheru emutauro wechiratidzo. Kudhirodha kwekufungidzira kunowanzo shandisa dhizaini diki kuratidza akati wandei tokeni uye yobva yabvunza hombe yemhando yemhando kuti ivaonese. Pepa rinopikisa kuti dhizaini iyi inogadzira dambudziko kune masisitimu emutauro wekuona: diki autoregressive drafter inogona kusakwanisa kugadzirisa chifananidzo padanho rega rega, kunyangwe ruzivo rwekuona runogona kuita kuti magwaro anotevera atarisike.
GLANCE inoshandura basa remugadziri. Sekureva kwaibva, musoro weblock-diffusion unoverenga iyo inotarirwa modhi yakatosanganiswa-yemutauro wemutauro uye inozadza chivharo chemumiriri mune imwechete yekupfuura. Muti wakafara wemumiriri wobva watariswa mune imwe chinangwa-modhiyo pass. Iyi ndiyo yepakati tekinoroji yekuda: ruzivo rwekuona rwatovepo mudunhu rakapihwa kune anodhirowa, nepo ma tokeni akarongwa anogadzirwa sechivharo kwete sequentially. Bepa rinodaidza nzira iyi isingarasikirwe nekuti yakanangana neiyo isina kugadziridzwa yemutauro-yemutauro modhi uye inoshuma kuti yega yega yakaongororwa inoburitsa kudhirodha kwemakaro chaizvo.
Vanyori vanorondedzera kuenzanisa kwakawanda pasi peyavanodaidza injini imwe uye imwe bhajeti. GLANCE decoded anosvika 2.93 nguva nekukurumidza kupfuura autoregressive decoding, yakashandisa imwe dhizaini pass pese pese apo kugadzirwa EAGLE3-VL musoro wakashandisa masere, uye yakagamuchirwa mabhuroko 2.7 nguva yakareba kupfuura EAGLE-3 musoro wakadzidziswa pane imwechete corpus. Iro bepa zvakare rinoshuma hukama pakati peyakagamuchirwa kureba kwebhuroka uye iyo inotarirwa modhi inotevera-chiratidzo entropy. Hukama ihwohwo hwakawedzera sezvo mabasa aivimba zvakanyanya nekuisa pasi, pamabasa mashanu, uye vanyori vanoti yakapfuura nepakati pezvinangwa uye modalities. Izvi zvichemo kubva kune preprint, kwete zvakasimbiswa zvakazvimirira.
Nei zvichikosha
Iro basa rinogadzirisa bhodhoro rinoshanda mune multimodal AI: kugadzira chinyorwa chimwe chete tokeni panguva inogona kuita kuti mhinduro dzemufananidzo dzidhure uye dzinonoka. Kana mhedzisiro yakashumwa ikawanda, GLANCE inogona kuderedza inference latency uye kuverengera mabasa anokopa kana kunyatsotevedzera zvinoonekwa zvemukati, uku ichichengetedza yaivavarirwa modhi inobuda pane kutengesa kurongeka kwekumhanya.
Iko kukosha kwekukurumidza ndeye computational. Mamodheru emutauro wechiratidzo anosanganisa kugadziridzwa kwemifananidzo nekugadzirwa kwemutauro, asi nhanho yechizvarwa inogona kuramba ichienderera nechiratidzo. Pamhinduro dzinodzokorodza mavara anoonekwa pamufananidzo kana kutevedzera diki diki rekuona, iyo nzira yekutevedzana inogona kuita nhanho dzakawanda dzinenge dzafungidzirika. Nzira inodhirowa kureba mukupasa imwe inogona kuderedza latency uye huwandu hwechinangwa-modhiyo computation inodiwa pamhinduro yega yega.
Musiyano wesosi pakati pezvinyorwa zvepasi uye zvemahara-zvinomhanya zvakakosha. GLANCE haina kuratidzwa sechitsividzo chepasi rose che autoregressive decoding. Vanyori vanoti mukana wayo wakanyanya kusimba mu "verbatim-copy regime," uko kutariswa kwekuona kunoita kuti kutevedzana kwakareba kufanofanotaurwa, nepo zvinyorwa zvemahara zvichiri kufarira cheni. Muganhu iwoyo unopa mhedzisiro chirevo chinoshanda: kuita kunogona kutsamira zvishoma pakuti modhi ine multimodal muuzhinji pane kuti ibasa rakadii rinomanikidza mhinduro kuburikidza nehumbowo hwekuona.
Kurasikirwa kunonzi kusarasikirwa kunogonawo kuve nebasa pamasarudzo ekutumira. Speedups inowanzoda kubvuma shanduko inogona kuitika mune zvinobuda, asi bepa iri rinoti mhedzisiro yemakaro yemodhi yakaburitswa pane zvayakaongororwa. Kana yakasimbiswa pabvunzo dzakakura, izvo zvinoita kuti hunyanzvi uhu huenderane kumashandisirwo anoda iyo yekutanga modhi yekudhirodha maitiro uchitsvaga yakaderera nguva yekupindura kana mutengo wekufungidzira. Iyo sosi haitsigire iyo yakadzika mari yekuchengetedza, zvisinei, uye haitauri kuendeswa kwekugadzirwa, kuwanikwa-kwakatarisana nemushandisi, kana mhedzisiro pamhando inodarika kuberekana muongororo yakaongororwa.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Iro bepa ichangobva kutumirwa preprint, saka zvarinoda zvichiri kuda kudzokororwa kwakazvimirira. Iyo abstract haitsanangure hardware, tarisiro modhi, dataset, kukurumidza kuverenga, absolute latency, kana iwo mashanu akaongororwa mabasa. Iyo yakashumwa yakanyanya kumhanyisa inogona saka kutsamira zvakanyanya pamutoro webasa uye sarudzo dzekuita; vanyori pachavo vanoti zvinyorwa zvinomhanya zvakasununguka zvichiri kufarira kunyora kwakatevedzana.
Mubvunzo mukuru ndeye reproducibility. Sosi yacho inozivisa bepa, vanyori, zuva rekutumira, uye kuwanikwa kwekodhi, asi chirevo chayo hachipi iyo yekushandisa URL, gadziriso yehardware, mazita emhando, saizi yedataset, kuverenga nekukurumidza, kana tsanangudzo dzebasa dzakadzama. Iwo madeti anodiwa kuti uone kana iyo 2.93 × nhamba inoratidza kuvandudzwa kwakakura kana mhedzisiro yakanakisa pane rimwe basa. Izwi rekuti "kusvika" rinoratidza kuti huwandu hwacho hausi hwakajairwa.
Kuenzanisa kunodawo kududzirwa kwokungwarira. GLANCE inofananidzwa neautoregressive decoding uye neEAGLE-yakavakirwa misoro pasi peinjini yakataurwa uye yakatenderedza bhajeti, asi iyo abstract haitsanangure kana masisitimu ese akashandisa akafanana kernels, batching, ndangariro marongero, kana maitirwo emitemo. Chirevo chekuti GLANCE yakagamuchira mabhuroko akapetwa ka2.7 kupfuura musoro we EAGLE-3 wakadzidziswa pane imwe corpus inodzidzisa, asi kureba kwakagamuchirwa chete hakuratidze kupera-kusvika-kumagumo latency. Mutengo wekusimbisa, kushandisa ndangariro, kushandiswa kwehardware, uye kutadza kana kuramba mitengo zvese zvine basa.
Kumwe kuongorora kunofanirwa kuyedza nzira pane mamwe mamodheru anonangwa, maitiro, mitauro, mhando dzemifananidzo, uye masitaera ekupindura, kusanganisira nyaya dzekuti mhinduro yakavhurika pane kukopa kubva pane yekuona. Hukama hwe entropy hunogona kubatsira nekuti hunogona kubatsira kufanotaura kuti kudhirowa kweblock kuchabhadhara riini, asi kwazvinopa hapana nguva yekusagadzikana kana kusimbiswa kwakazvimirira kwemutemo wakatsanangurwa. Kusvika zviyero izvi zvavepo, GLANCE inonzwisiswa zvakanyanya seyakavimbiswa tsvakiridzo kubva kune preprint pane yakaratidza general-chinangwa chekuwedzera mwero.