I-Google imemezele isilingo sokushumeka imodeli yolimi lomshini olukhulu ku-Chrome. Ukuze ufinyelele imodeli kusuka kuzinhlelo zokusebenza zewebhu nezengezo zesiphequluli, i-Prompt API iyahlongozwa, ekuvumela ukuthi uthumele izicelo ngolimi lwemvelo, olufana nama-chatbots. Kulindeleke ukuthi imodeli yolimi enkulu eyakhelwe esipheqululini izokwenza kube lula ukwenziwa kwemisebenzi ye-AI ezinhlelweni zewebhu futhi izoqeda isidingo sokukhathazeka ngokufaka nokuphatha amamodeli olimi.
Ukuhlolwa okusetshenzisiwe imodeli ye-Gemini Nano, ehlangene kakhulu yomndeni wakwaGemini. Kungenzeka futhi ukufaka amamodeli ochwepheshe anweba imodeli eyisisekelo ngolwazi olwengeziwe olungadingeka ukuze kuxazululwe izinkinga ezithile, kanye nokusebenzisa amakhono afana nokuhumusha komshini nokufingqa. Amamodeli asetshenziswa endaweni ohlelweni lomsebenzisi ngaphandle kokufinyelela izinsiza zangaphandle.
Isikhathi sokusebenza esisetshenziselwa ukusebenzisa imodeli sisebenzisa ngokuzenzakalelayo ama-GPU nama-NPU atholakala kusistimu ukusheshisa umsebenzi ngemodeli noma ukushintshela ekusebenziseni imodeli kusetshenziswa i-CPU. Izinzuzo zokusebenzisa imodeli kusistimu yomsebenzisi zihlanganisa ukugcinwa kuyimfihlo kwedatha ecutshunguliwe, ikhono lokuqhubeka nokusebenza kumodi engaxhunyiwe ku-inthanethi ngaphandle kokuxhumeka kwenethiwekhi noma uma izinkinga ziphakama ngekhwalithi yokuxhumana, ukunciphisa ukubambezeleka lapho kuthunyelwa izicelo, kanye nokuqeda ukuncika ezinsizeni zangaphandle.
I-API Prompt, eyenzelwe ukusebenzisana nemodeli, ikuvumela ukuthi wenze izicelo ezilula kuphela ngolimi lwemvelo, kodwa futhi uhlele ukubandakanyeka kwemodeli ekucubunguleni nasekuhlukaniseni idatha ngokucabangela umongo, ucabangele izicelo nedatha ngaphambilini. ithunyelwe kuseshini, futhi usebenzise imodeli ukuze ukhethe izinketho ezingcono kakhulu (ngokwesibonelo, ungacela ukukhetha isithonjana ohlwini lwama-emoji ukuze uthole amazwana athile kusayithi). Ukwengeza, kuhlelwe ukuthuthukisa i-API ukuze isetshenziswe emkhakheni wokudala okuqukethwe nokuxazulula izinkinga ezinjengokubhala kabusha ngamanye amagama, ukuhlola nokulungisa uhlelo lolimi.
Ngokuvamile, izinhlobo ezimbili ze-API ziyathuthukiswa ukuze zihlanganyele nemodeli ye-AI eyakhelwe ngaphakathi - Umsebenzi Nokuhlola. Esokuqala sinikeza ukufinyelela kumakhono okuxazulula izinkinga ezithile, njengokuhumusha umbhalo usuka kolunye ulimi uye kolunye (API Translation) noma ukufingqa ingqikithi eyinhloko yombhalo (API Summarization). Uhlobo lwesibili luhloselwe ukudala nokuhlola ama-prototypes okuhlola lapho kwakhiwa ama-Task API amasha. Umsebenzi futhi uyaqhubeka nge-LoRA (Low-Rank Adaptation) API ukulungisa ama-coefficients okulinganisa wemodeli yesisekelo ukuze kuthuthukiswe ukusebenza kahle kokuxazulula izinkinga ezithile.
Ukufinyelela kokubamba iqhaza ekuhlolweni kunikezwa ngemva kokugcwalisa isicelo. I-API ingaphansi kokuthuthukiswa okusebenzayo futhi izonwetshwa futhi ishintshwe kucatshangelwa impendulo yomsebenzisi kanye nezintandokazi kuze kube inguqulo yokugcina yamukelwe. Ngokuzayo, sihlela ukuhlela ukuhlola okufinyeleleka kakhudlwana sisebenzisa imodi ye-Origin Trials, enikeza amandla okusebenza nama-API okuhlola kusukela kuzinhlelo zokusebenza ezilandwe ku-localhost noma i-127.0.0.1, noma ngemva kokubhalisa nokuthola ithokheni elikhethekile elivumeleke isikhathi esilinganiselwe. isikhathi sesayithi elithile. Ngokuhambisana nabakhiqizi bezinye iziphequluli, umsebenzi uyaqhubeka wokulinganisa ama-API athuthukisiwe.
Source: opennet.ru
