I-FlexGen iyinjini yokusebenzisa i-ChatGPT-like AI bots kumasistimu e-GPU eyodwa

Ithimba labacwaningi abavela eNyuvesi yaseStanford, iNyuvesi yaseCalifornia eBerkeley, i-ETH Zurich, i-Graduate School of Economics, i-Carnegie Mellon University, kanye ne-Yandex ne-Meta, bashicilele ikhodi yomthombo wenjini yokusebenzisa amamodeli ezilimi ezinkulu esisetshenziswa. -izinhlelo eziphoqelekile. Isibonelo, injini inikeza ikhono lokudala ukusebenza okukhumbuza i-ChatGPT kanye ne-Copilot ngokusebenzisa imodeli ye-OPT-175B eqeqeshwe ngaphambilini, ehlanganisa amapharamitha ayizigidi eziyizinkulungwane ezingu-175, kukhompuyutha evamile enekhadi lemifanekiso yegeyimu ye-NVIDIA RTX3090 efakwe u-24GB wememori yevidiyo. Ikhodi ibhalwe nge-Python, isebenzisa uhlaka lwe-PyTorch futhi isatshalaliswa ngaphansi kwelayisensi ye-Apache 2.0.

Kuhlanganisa nombhalo oyisibonelo wokudala ama-bots okuvumela ukuthi udawunilode eyodwa yamamodeli olimi atholakala esidlangalaleni bese uqala ukuxhumana ngokushesha (ngokwesibonelo, ngokusebenzisa umyalo othi “python apps/chatbot.py —model facebook/opt-30b — -percent 0 100 100 0 100 0”). Njengesisekelo, kuhlongozwa ukuba kusetshenziswe imodeli yolimi enkulu eshicilelwe yi-Facebook, eqeqeshwe ngamaqoqo e-BookCorpus (izincwadi eziyizinkulungwane ezingu-10), i-CC-Stories, i-Pile (OpenSubtitles, i-Wikipedia, i-DM Mathematics, i-HackerNews, njll.), i-Pushshift. io (kusekelwe kudatha ye-Reddit ) kanye ne-CCNewsV2 (ingobo yomlando yezindaba). Imodeli ihlanganisa cishe amathokheni ayizigidi eziyizinkulungwane ezingu-180 (800 GB idatha). Izinsuku ezingama-33 zokusebenza kweqoqo elinama-992 NVIDIA A100 80GB GPUs zachithwa ekuqeqesheni imodeli.

Lapho isebenzisa imodeli ye-OPT-175B kusistimu ene-NVIDIA T4 GPU (16GB) eyodwa, injini ye-FlexGen ibonise ukusebenza ngokushesha okufika izikhathi eziyi-100 kunezixazululo ezazinikelwe ngaphambili, okwenza ukusetshenziswa kwamamodeli wezilimi ezinkulu kufinyeleleke kakhudlwana futhi kuzivumela ukuthi ziqhubeke zisebenza. amasistimu ngaphandle kwama-accelerator azinikele. Ngesikhathi esifanayo, i-FlexGen ingakala ukuze ihambisane nezibalo ngama-GPU amaningi. Ukuze kuncishiswe usayizi wemodeli, isikimu sokuminyanisa ipharamitha yobunikazi kanye nendlela yokugcinwa kwesikhashana eyimodeli kuyasetshenziswa.

Njengamanje, i-FlexGen isekela kuphela amamodeli wolimi we-OPT, kodwa esikhathini esizayo abathuthukisi bathembisa ukwengeza ukwesekwa kwe-BLOOM (amapharamitha ayizigidi eziyizinkulungwane ezingu-176, isekela izilimi ezingu-46 nezilimi ezingu-13 zokuhlela), i-CodeGen (ingakwazi ukukhiqiza ikhodi ngezilimi zokuhlela ezingu-22) futhi Amamodeli we-GLM. Isibonelo sengxoxo ne-bot esekelwe ku-FlexGen kanye nemodeli ye-OPT-30B:

Umuntu: Ithini igama lentaba ende kunazo zonke emhlabeni?

Umsizi: Everest.

Buntu: Ngihlela uhambo lokugubha usuku lwethu. Yiziphi izinto esingayenza?

Umsizi: Nokho, kunezinto ezimbalwa ongazenza ngosuku lwakho lokugubha usuku. Okokuqala, ungadlala amakhadi. Okwesibili, ungahamba uyoshaywa umoya. Okwesithathu, ungaya emnyuziyamu.

Source: opennet.ru