Kwenzekeni
I-Apple Ucwaningo Lokufunda Ngomshini lushicilele iphepha elichaza i-STARFlow2, isakhiwo esikhiqiza izinhlobo eziningi esakhelwe ukuqonda, ukucabangisisa futhi sikhiqize ukulandelana kwezithombe zombhalo ezishiyekile. Ababhali bathi uhlelo lusebenzisa ukugeleza okuzenzakalelayo kwe-autoregressive normalizing eduze kwemodeli yolimi lombono oluqeqeshelwe kusengaphambili, ngayo yomibili imifudlana esebenza ngaphansi kwesakhiwo esifanayo sembangela. I-Apple ibika imiphumela enamandla kuwo wonke ama-benchmarks esizukulwaneni sesithombe kanye ne-multimodal, kodwa umthombo awunikezi amagama ebhentshimakhi, amaphuzu noma iziqhathaniso.
Ikhasi le-Apple Machine Learning Research, elimakwe njengelishicilelwe ngo-Agasti 2026, lethula i-STARFlow2 njengocwaningo ekwenziweni okuhlanganisiwe kwe-multimodal. Iphepha likhuluma ngezinhlelo ezingacubungula futhi zikhiqize ukulandelana lapho umbhalo nezithombe zishiyana. Ababhali be-Apple baphikisa ngokuthi izindlela zamanje zihlukaniswe ngokwesakhiwo: ezinye zithembele ekwazisweni okubonakalayo okunganciphisa ukwethembeka okubonakalayo, abanye bahlanganisa isizukulwane solimi oluyimbangela nokuphindwaphindwa kokusabalalisa okuphindaphindekayo, futhi amanye avumelanise amamodeli olimi lombono ukuze akhiqize ngezindlela ezingase zenze buthaka ukuqonda kwazo okulungiselelwe kusengaphambili. Lezi yizimpawu zephepha zezindlela ezikhona, hhayi okutholwe ngokuzimela okunikezwe yilo mthombo.
Isiphakamiso esimaphakathi siwukusebenzisa ukugeleza kwe-autoregressive normalizing njengohlaka olujwayelekile olukhiqizayo lolimi nokuphumayo okubukwayo. Isakhiwo esibikiwe sakhelwe kulokho iphepha elikubiza ngokuthi i-Pretzel design. Isuka iqonde phakathi kwemifudlana emibili: ukusakaza okuqandisiwe okuqandisiwe kwemodeli yolimi kanye nokusakaza kwe-TARFlow. Izixhumanisi zokweqa eziyinsalela zixhuma ukusakaza, futhi zombili zisebenza ngaphansi kwemaski yembangela efanayo. Umthombo uchaza ukugeleza kwe-autoregressive normalizing flows njengama-autoregressive Transformers abelana nge-causal mask, indlela yenqolobane yokhiye-ukhiye kanye nesakhiwo esisuka kwesokunxele siye kwesokudla esinamamodeli amakhulu olimi.
I-STARFlow2 iphinda ihlanganise idizayini yokugeleza engajulile nesikhala esifihlekile se-FAE. I-Apple ithi lokhu kuvumela isistimu ukuthi ikhiqize okuqukethwe okunqamukile ngendlela evumelana nenqolobane, ngombhalo nokuphumayo okubukwayo kungena kunqolobane yenani lokhiye ngokuqondile esikhundleni sokubhalwa kabusha kwekhodi. Ikhasi lithi ukuhlolwa kubonisa ukusebenza okuqinile kuwo wonke amabhentshimakhi esizukulwane sesithombe kanye ne-multimodal-understanding, i-Apple eyethula njengokuqinisekisa ukuthi ukugeleza okuzenzakalelayo kungasebenza njengesisekelo sokumodela okuhlanganisiwe kwe-multimodal. Nokho, umthombo onikeziwe awuwahlonzi amabhentshimakhi, amaphuzu wombiko, amasistimu wokuqhathanisa amagama, uchaza amasethi edatha noma usho izindleko zokubala. Futhi ayisho ukuthi imodeli, ikhodi, izisindo noma ukuboniswa okusebenzisanayo kuyatholakala. Ikhasi libala imisebenzi ehlobene ekwenziweni kwevidiyo ngokugeleza okujwayelekile, kodwa lowo msebenzi uhlukile kusimemezelo se-STARFlow2. Ebufakazini obunikeziwe, lona umphumela wocwaningo kanye nesiphakamiso sezakhiwo, hhayi ukwethulwa komkhiqizo noma isevisi yomphakathi ebonisiwe.
Imininingwane yomthombo: machinelearning.apple.com ↗
Kungani kubalulekile
Umsebenzi uqondise inkinga yesakhiwo ku-AI ye-multimodal: amamodeli olimi ngokuvamile akhiqiza amathokheni ahlukene, kuyilapho izithombe ziyidatha eqhubekayo futhi ngokuvamile zikhiqizwa ngamasistimu ahlukene okusabalalisa noma aphindaphindayo. Uma idizayini ebikiwe ibambelela, indlela eyodwa eyimbangela ingase yenze kube lula ukukhiqizwa kwe-multimodal futhi yenze amasistimu ezithombe zombhalo ahlukene asebenze kalula. Inani elingokoqobo lihlala lingaqinisekisiwe ngoba umthombo awunikezi ukubambezeleka, izindleko, ikhwalithi noma idatha yokutholakala.
Ukubaluleka okungokoqobo kwe-STARFlow2 umzamo wayo wokuxazulula ukungafani phakathi kolimi nokukhiqizwa okubukwayo ngaphakathi kwesakhiwo semodeli eyodwa. Ku-akhawunti yephepha, ukukhiqizwa kolimi ngokwemvelo kusingathwa njengenqubo esuka kwesokunxele uye kwesokudla, kuyilapho ukukhiqizwa kwesithombe kuvame ukusingathwa ngokumelela okuhlukile noma inqubo yokuphindaphinda umsindo. Indlela ye-Apple ehlongozwayo esekelwe ekugelezeni ihloselwe ukugcina zombili izindlela ngaphakathi kokulandelana okuqhubekayo, okuyimbangela. Uma ukuhlola okuzimele kuqinisekisa isimangalo, lokhu kungase kunikeze abacwaningi idizayini ehambisana kakhudlwana yezinhlelo zokusebenza ezidinga ukushintshanisa phakathi kokuchaza isithombe, ukusikhiqiza, ukusichaza nokuqhubeka nombhalo.
Idizayini yenqolobane ingomunye umnikelo ongase ube wusizo. I-Apple ithi okuphumayo kombhalo nokubonwayo kungangena ngokuqondile kunqolobane yenani lokhiye ngaphandle kokufaka ikhodi kabusha. Empeleni, lokho kungase kunciphise ukucutshungulwa okuyimpinda lapho isistimu ihamba ngokuphindaphindiwe phakathi kwezindlela, ikakhulukazi ezingxoxweni ezinde ezinamakhefu noma imisebenzi yesizukulwane. Umthombo awuwubali umphumela, nokho. Azikho izilinganiso zokubambezeleka okubikiwe, izibalo zenkumbulo, imiphumela yokuphuma, izindleko zokuphakela noma ukuqhathanisa nesistimu esekelwe ekusakazweni noma yamathokheni efanayo. Ngakho-ke inzuzo efunwayo kufanele ithathwe njengenhloso yezakhiwo kanye nesimangalo socwaningo esikhundleni senzuzo yokusebenza emisiwe.
Lesi siphakamiso sibalulekile futhi ngoba sizama ukulondoloza ukuqonda kwemodeli eqeqeshelwe kusengaphambili ngenkathi yengeza ukwakhiwa kwesithombe esithembekile kakhulu. I-Apple ithi ukusakaza okufriziwe kolimi lombono kusiza ukugcina ukuqonda okukhona nokuthi ukusakaza kunikeza amandla ukukhiqiza okubonakalayo. Leyo nhlanganisela ingase ihlobane kumasistimu esikhathi esizayo adinga kokubili ukuhunyushwa nokudalwa, kodwa umthombo awuqinisekisi ukuthi kungakanani ukuqonda okulondoloziwe, ukuthi ukwethembeka okubonakalayo kukalwe kanjani noma ukuthi izinzuzo ngomsebenzi owodwa ziyahwebelana ngokuqhathaniswa nokusebenza komunye. Imiphumela yephepha ingase ibaluleke ocwaningweni lwamamodeli amaningi, nokho ukubaluleka kwawo okubanzi kuncike emininingwaneni engekho esimemezelweni.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ubufakazi obubalulekile obulandelayo ukushicilelwa okugcwele, okuhlanganisa imininingwane yebhentshimakhi, izisekelo, ukukhishwa, usayizi wamamodeli, izidingo zokubala kanye nanoma iyiphi ikhodi noma izisindo. Abacwaningi nabathuthukisi kufanele futhi bahlole ukuthi izinzuzo ze-KV-cache ezifunwayo ziyaqhubeka yini nokulandelana okude okunezinhlangothi nokugeleza komsebenzi kwangempela kwe-multimodal. I-Apple ayishongo kulo mthombo ukuthi i-STARFlow2 iwumkhiqizo, imodeli yomphakathi noma ithuluzi elitholakala ngokuvamile, futhi usuku oluqondile lokushicilelwa phakathi kuka-Agasti 2026 alunikezwanga.
Iphuzu lokuqala lokuqinisekisa yiphepha eliphelele kanye nerekhodi lalo lokuhlola. Ubufakazi obuwusizo bokulandelela buzohlanganisa amagama nezinguqulo zamabhentshimakhi esizukulwane sesithombe kanye nokuqonda okuningiliziwe, imiphumela yezinombolo ye-STARFlow2 nezisekelo ezifanele, izifundo zokukhishwa ezihlukanisa ukusakaza kolimi lombono oluqandisiwe, ukusakaza kwe-TARFlow, ukuxhumeka okuyinsalela kanye nedizayini yesikhala esicashile, kanye nemininingwane mayelana nedatha yokuqeqeshwa, usayizi wemodeli nokubala. Ngaphandle kwaleyo mininingwane, "ukusebenza okuqinile" kubanzi kakhulu ukuthola ukuthi kukhulu noma kuthembekile kangakanani ukuthuthukiswa.
Umbuzo wesibili ukuthi ingabe isakhiwo esihlongozwayo sembangela kanye ne-cache-friendly sisebenza ngaphezu kokuhlolwa okubikiwe. Abacwaningi abazimele bazodinga ukuhlola ukulandelana okude okuqukethe ukuguqulwa okuphindaphindiwe phakathi kombhalo nezithombe, njengoba ukukhula kwenqolobane, ukusetshenziswa kwenkumbulo nokunqwabelana kwamaphutha kungase kuthinte ukusebenza okungokoqobo. Kufanele futhi baqhathanise ikhwalithi yokuphumayo, ukungaguquguquki, ukulawuleka kanye nezindleko zokuhlinzeka ngokuqhathaniswa namasistimu kusetshenziswa amathokheni esithombe ahlukene noma ukuphindaphindeka kokusabalalisa. Umthombo awunikezi bufakazi okwamanje mayelana nalokhu kuhwebelana, futhi awubiki ukuziphatha ngaphansi kwamashifu okusabalalisa noma ukwaziswa okunzima kwe-multimodal.
Umbuzo wokugcina ukuthunyelwa. Umthombo awusho ukuthi i-STARFlow2 itholakala ngomkhiqizo we-Apple, i-API, inqolobane yocwaningo noma indawo yokuhlola elandekayo. Futhi ayinikezi usuku oluqondile lokushicilela, ngakho-ke isikhathi esiphakathi kwewindi lezindaba lamahora angu-96 asikwazi ukuncishiswa ngokuqhubekayo ezintweni ezinikeziwe. Ukubika kokulandelela kufanele kuveze ukuthi i-Apple iyayikhipha yini imininingwane yokusetshenziswa noma ama-artifact angamamodeli, ukuthi amaqembu angaphandle angakwazi yini ukukhiqiza kabusha imiphumela, kanye nokuthi izakhiwo ziyawathonya yini amasistimu ezinto eziningi ezentengiselwano. Kuze kube yileso sikhathi, okubalulekile okungaziwa ukutholakala, ukukhiqizwa kabusha, izidingo zensiza kanye nezinzuzo ezikalwa ngokuzimela.