Imodeli ye-AI evulekile ye-Hertz-dev yokuxhumana kwezwi okuphindwe kabili okushicilelwe

I-Standard Intelligence imemezele ukushicilelwa kwe-hertz-dev, imodeli yokuqala ye-AI evulekile yokuhlanganisa inkulumo egcwele i-dupex engasetshenziswa njengesisekelo sokwakha ukuxhumana kwezwi ngesikhathi sangempela noma izinhlelo zokukhiqiza ulimi olukhulunywayo. Imodeli ingakwazi ukukhiqiza inkulumo efana kakhulu nedatha yezwi lapho iqeqeshelwa khona, inikeze ukusebenzelana kwesitayela somuntu ngaphandle kokushiyeka kwengxoxo yocingo. Intuthuko yephrojekthi isatshalaliswa ngaphansi kwelayisensi ye-Apache 2.0.

Kusistimu ene-NVIDIA GeForce RTX 4090 GPU, ukubambezeleka okumaphakathi kwesizukulwane sangaphambi kwesikhathi ngu-120 ms (ngokweqile kufika ku-65 ms), okushesha cishe ngokuphindwe kabili kunamamodeli akhona atholakala esidlangalaleni. Inguqulo eshicilelwe yakhiwe kusetshenziswa i-Transformer architecture, ihlanganisa imingcele ye-8.5 billion futhi iqeqeshwa ngokusebenzisa amathokheni ayizigidi eziyizinkulungwane ezingu-500. Usayizi womongo ocatshangelwe yimodeli (inani lamathokheni imodeli engakwazi ukuwacubungula futhi uwakhumbule lapho ikhiqiza inkulumo) amathokheni angu-2048 noma cishe imizuzu emi-4 yokukhuluma.

Source: opennet.ru

Thenga ukusingathwa okuthembekile kwamasayithi anokuvikelwa kwe-DDoS, amaseva e-VPS VDS 🔥 Thenga ukusingathwa kwewebhusayithi okuthembekile ngokuvikelwa kwe-DDoS, amaseva e-VPS VDS | ProHoster