PrismML Isibuyekezo: Ikusasa Lamamodeli Olimi Amakhulu Asebenza Kumadivayisi

#Quasa #QUA #prismml

I-PrismML iyinkampani yocwaningo lwe-AI ezinze e-U.S. (enezimpande eziqinile e-Caltech) egxile enseleleni ebalulekile: indlela yokwandisa ubuhlakani bamamodeli ngaphandle kokwandisa ubukhulu bawo noma ubunkimbinkimbi bawo. Isixazululo sayo ngamamodeli aminyene kakhulu nacindezelwe kakhulu asebenza kahle kuyilapho esebenzisa inkumbulo, amandla nezinsiza zokucubungula okuncane kakhulu.

Umkhiqizo wayo ohamba phambili, i-Bonsai 27B, uyimodeli enamapharamitha ayizigidigidi ezingu-27 etholakala ngezinguqulo ze-1-bit (engu-3.9 GB kuphela) neze-ternary (ezingu-5.9 GB). Lokhu kuvumela imodeli yalobu bukhulu ukuba isebenze ngokuqondile kuma-smartphone nakumakhompyutha aphathekayo — into eyayingenakwenzeka ngaphambili ngamamodeli avamile asebenzisa ukunemba okuphelele. I-PrismML ithi inguqulo ye-1-bit isebenzisa inkumbulo encane ngokuphindwe ngo-14×, isebenza ngokushesha ngokuphindwe ngo-8×, futhi isebenzisa amandla amancane ngokuphindwe ngo-5× uma iqhathaniswa namamodeli endabuko e-16-bit, kuyilapho isasekela ukucabanga ngezinyathelo eziningi, ukubiza amathuluzi, ukugeleza komsebenzi okwenziwa ama-ejenti e-AI, nokuqonda izinhlobo eziningi zokwaziswa.

Ifilosofi eyisisekelo yenkampani “ukuminyana kobuhlakani” — ukwandisa ngangokunokwenzeka ubuhlakani obuwusizo nge-bit ngayinye kunokumane kwandiswe inani lamapharamitha. Le ndlela ihlose ukuxazulula izinkinga ezimbili ezinkulu ngesikhathi esisodwa: amamodeli amakhulu angangeni kumadivayisi omuntu siqu, kanye nezidingo zamandla ezingasimame zezikhungo ezinkulukazi zedatha.

I-PrismML ifanele abathuthukisi, abakhiqizi bezingxenyekazi zekhompyutha, abacwaningi nezinkampani ezifuna i-AI enamandla esebenza khona kudivayisi, enokuvikeleka okuqinile kobumfihlo, ukubambezeleka okuncane nokusebenzisa amandla kahle kakhulu.

Amaphuzu Avelele

• Ukucindezela kwe-1-bit nokwe-ternary okusezingeni lomhlaba (umndeni we-Bonsai).
• I-Bonsai 27B isebenza kumafoni isebenzisa i-~4 GB kuphela.
• Intuthuko enkulu ekusetshenzisweni kwenkumbulo, esivinini nasekusetshenzisweni kwamandla kahle.
• Isekela ukucabanga, ukusetshenziswa kwamathuluzi namakhono okusingatha izinhlobo eziningi zokwaziswa khona kudivayisi.
• Isisekelo esiqinile socwaningo esivela e-Caltech.

Izinto Ezingase Zidinge Ukucatshangelwa

• Njengoba lobu buchwepheshe busasha uma buqhathaniswa nobunye, ukusekelwa ohlelweni lobuchwepheshe obuhambisanayo namathuluzi okucija amamodeli kusaqhubeka nokwanda.
• Okwamanje imiphumela emihle kakhulu ibonakala ezingxenyekazini zekhompyutha zesimanje ezisezingeni eliphezulu zamadivayisi eselula namakhompyutha aphathekayo.

Isinqumo Esiphelele: izinkanyezi ezingu-4.6/5

I-PrismML ithuthukisa omunye wemikhakha ebaluleke kakhulu ye-AI njengamanje — yenza amamodeli amakhulu asebenziseke kumadivayisi asetshenziswa nsuku zonke. Ngokugxila ekuminyaneni kobuhlakani esikhundleni sobukhulu nje kuphela, inkampani isiza ukususa i-AI enamandla efwini iyiyise kumadivayisi aseduze nendawo yokusetshenziswa ngendlela enomthelela obonakalayo.

Nawe zuza umklomelo we-QUA ngokubhala isibuyekezo ku-Quasa.io!

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