PrismML: Ndemanga pa Tsogolo la Ma Model Akuluakulu a Chilankhulo Ogwira Ntchito pa Zipangizo

#Quasa #QUA #prismml

PrismML ndi kampani yofufuza za AI yomwe ili ku U.S. (yomwe inayambira kwambiri ku Caltech), ndipo ikuyang'ana kwambiri vuto lofunika kwambiri: momwe ingachulukitsire luntha la ma model popanda kuwonjezera kukula kwawo kapena kucholowana kwawo. Yankho lawo ndi ma model okhala ndi luntha lochuluka kwambiri poyerekezera ndi kukula kwawo, omwe achepetsedwa kwambiri kukula ndipo amagwira ntchito bwino kwambiri pamene amagwiritsa ntchito memori, mphamvu ndi mphamvu zowerengera zochepa kwambiri.

Chinthu chawo chachikulu, Bonsai 27B, ndi model yokhala ndi ma parameter 27 biliyoni, yomwe imapezeka m'mitundu ya 1-bit (3.9 GB yokha) ndi ternary (5.9 GB). Izi zimathandiza kuti model yaikulu chonchi igwire ntchito mwachindunji pa mafoni anzeru ndi ma laputopu — zomwe poyamba zinali zosatheka ndi ma model anthawi zonse osachepetsedwa kulondola kwa manambala awo. PrismML imati mtundu wa 1-bit umagwiritsa ntchito memori yocheperapo ndi 14×, umagwira ntchito mofulumira ndi 8×, ndipo umagwiritsa ntchito mphamvu zocheperapo ndi 5× poyerekezera ndi ma model achikhalidwe a 16-bit, uku ukuthandizabe kuganiza motsatira masitepe angapo, kuyitanitsa zida, njira zogwirira ntchito za ma agent, ndi kumvetsetsa mitundu yosiyanasiyana ya uthenga.

Mfundo yaikulu yomwe imatsogolera kampaniyi ndi “kuchuluka kwa luntha pa biti” — kuwonjezera kwambiri luntha lothandiza pa biti iliyonse m'malo mongowonjezera chiwerengero cha ma parameter. Njira imeneyi ikufuna kuthetsa mavuto awiri akuluakulu nthawi imodzi: ma model akuluakulu omwe sangakwane pa zipangizo za munthu payekha, ndi kuchuluka kwa mphamvu zomwe malo akuluakulu osungira ndi kukonza deta amafuna, komwe sikungapitirire kwa nthawi yaitali.

PrismML ndi yabwino kwa opanga mapulogalamu, opanga zipangizo, ofufuza, ndi makampani omwe akufuna AI yamphamvu yogwira ntchito pa zipangizo zawo, yosunga kwambiri chinsinsi cha deta, yoyankha mwachangu, komanso yogwiritsa ntchito mphamvu mosamala kwambiri.

Mfundo Zazikulu

• Kuchepetsa kukula kwa ma model pogwiritsa ntchito 1-bit ndi ternary pamlingo wapamwamba kwambiri padziko lonse (gulu la Bonsai).
• Bonsai 27B imagwira ntchito pa mafoni pogwiritsa ntchito ~4 GB yokha.
• Kusintha kwakukulu pa kagwiritsidwe ntchito ka memori, liwiro, ndi kagwiritsidwe ntchito kabwino ka mphamvu.
• Imathandiza kuganiza, kugwiritsa ntchito zida, ndi kugwira ntchito ndi mitundu yosiyanasiyana ya uthenga pa chipangizocho.
• Maziko olimba a kafukufuku ochokera ku Caltech.

Zinthu Zomwe Mungafunike Kuziganizira

• Popeza umenewu ndi ukadaulo womwe udakali watsopano, thandizo lochokera ku mapulogalamu ndi ntchito zogwirizana nawo komanso zida zosinthira ma model kuti agwire bwino ntchito inayake zikupitirizabe kukula.
• Panopa zotsatira zabwino kwambiri zikuoneka pa mafoni ndi ma laputopu amakono apamwamba kwambiri.

Chigamulo Chonse: Nyenyezi 4.6/5

PrismML ikupititsa patsogolo imodzi mwa mbali zofunika kwambiri za AI panopa — kuchititsa kuti ma model akuluakulu azitha kugwiritsidwa ntchito pa zipangizo za tsiku ndi tsiku. Poyang'ana kwambiri kuchuluka kwa luntha pa biti m'malo mongoyang'ana kukula, kampaniyi ikuthandiza kwambiri kusamutsa AI yamphamvu kuchokera ku cloud kupita ku zipangizo zomwe zili pafupi ndi ogwiritsa ntchito.

Pezaninso mphotho ya QUA polemba ndemanga pa Quasa.io!

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