PrismML Ihlolwa: Ikusasa Lamamodeli Amakhulu Olimi Asebenza Ematshinini

#Quasa #QUA #prismml

I-PrismML yinkampani yase-U.S. ecwaninga nge-AI (elezisekelo eziqinileyo ezivela e-Caltech) egxile enseleleni eqakathekileyo: indlela yokwandisa ubuhlakani bamamodeli ngaphandle kokwandisa ubukhulu bawo kumbe ukwenza ukusebenza kwawo kube nzima kakhulu. Isixazululo sabo ngamamodeli aminyaniswe kakhulu, alobuhlakani obunengi endaweni encane, asebenza kuhle kakhulu esebenzisa inkumbulo, amandla kagesi lezinsiza zokubala okulutshwana kakhulu.

Umkhiqizo wabo ohamba phambili, Bonsai 27B, yimodeli elamapharamitha angamabhiliyoni angu-27 etholakala ngezinguqulo ze-1-bit (eyi-3.9 GB kuphela) leze-ternary (eyi-5.9 GB). Lokhu kwenza imodeli enkulu kangaka isebenze ngokuqondileyo kuma-smartphone lakuma-laptop — into ebingeke yenzeke ngaphambilini ngamamodeli avamileyo asebenzisa ukunemba okupheleleyo. I-PrismML ithi inguqulo ye-1-bit isebenzisa inkumbulo ephindwe ka-14× ukuba ncane, isebenza ngesiqubu esiphindwe ka-8×, njalo isebenzisa amandla kagesi aphindwe ka-5× ukuba malutshwana nxa iqathaniswa lamamodeli endabuko e-16-bit, ilokhu isekela ukucabanga ngezinyathelo ezinengi, ukubiza amathuluzi, ukuhleleka kwemisebenzi eyenziwa ngama-ejenti e-AI, lokuqonda izinhlobo ezitshiyeneyo zolwazi.

Umqondo oyisisekelo senkampani nguwo “ubunengi bobuhlakani endaweni encane” — ukwandisa ubuhlakani obulusizo kubhithi ngalinye kulokumane kwandiswe inani lamapharamitha. Indlela le ihlose ukuxazulula inhlupho ezimbili ezinkulu ngasikhathi sinye: amamodeli amakhulu angeneli ematshinini omuntu siqu, lezidingo zamandla kagesi ezinkulu kakhulu zezikhungo zedatha ezinkulukazi ezingeke ziqhubeke zifezwa isikhathi eside.

I-PrismML ilungele abathuthukisi bezinhlelo, abenzi bezingxenye zekhompuyutha, abacwaningi, lezinkampani ezifuna i-AI elamandla esebenza ematshinini azo, ilokuvikeleka okuqinileyo kolwazi lwangasese, ukuphuza okulutshwana ekuphenduleni, lokusebenzisa amandla kagesi ngendlela eyonga kakhulu.

Okuqakathekileyo

• Ukuminyanisa nge-1-bit lange-ternary okusezingeni eliphezulu emhlabeni (iqembu lamamodeli e-Bonsai).
• I-Bonsai 27B isebenza emafonini isebenzisa i-~4 GB kuphela.
• Ukuthuthuka okukhulu ekusetshenzisweni kwenkumbulo, esiqubwini, lasekongeni amandla kagesi.
• Isekela ukucabanga, ukusebenzisa amathuluzi, lamakhono okusebenzisa izinhlobo ezitshiyeneyo zolwazi emtshinini ngokwawo.
• Isisekelo esiqinileyo socwaningo esivela e-Caltech.

Okungadinga Ukucatshangelwa

• Njengoba lobu kuyibuchwepheshe obusesebutsha, ukusekelwa yizinhlelo lamathuluzi asebenzisana labo, kanye lamathuluzi okulungisa amamodeli ngokuqondileyo ukuze enze imisebenzi ethile, kulokhu kusanda.
• Khathesi imiphumela emihle kakhulu ibonakala ezingxenyeni zama-smartphone lama-laptop zesimanje ezisezingeni eliphezulu.

Isinqumo Sisonke: izinkanyezi ezingu-4.6/5

I-PrismML iqhubela phambili enye yezindawo eziqakatheke kakhulu ekuthuthukisweni kwe-AI khathesi — ukwenza amamodeli amakhulu asebenziseke ematshinini asetshenziswa nsuku zonke. Ngokugxila ebunengini bobuhlakani endaweni encane kulokugxila ebukhulwini nje, inkampani incedisa ukususa i-AI elamandla ezinhlelweni zamayezi iyise ematshinini aseduze labasebenzisi ngendlela elomthelela oqakathekileyo.

Thola umvuzo we-QUA ngokubhala ukuhlola ku-Quasa.io lawe!

Qalisa: https://quasa.io/projects/prismml