Yeqa uye kokuyinhloko
Ama-modeli kanye nobuchwepheshe

Ukuhambisana kwe-Day0: Iseva ye-Aliyun Lingjun Zhenwu M890 supernode manje isekela imodeli enkulu i-Kimi K3 ye-Moonshot

I-Aliyun imemezele ukuthi iseva yayo ye-Lingjun Zhenwu M890 supernode isihambisana ngempumelelo nemodeli yakamuva ephambili ye-Moonshot, i-Kimi K3, enamapharamitha angu-2.8 trillion. Ngokusekelwe kuma-chip e-AI e-T-Head ahlanganisa ukuqeqesha nokusebenzisa kanye nesakhiwo sokuxhumana ngesivinini esikhulu, le seva ingagcina ithrafikhi yokuxhumana kochwepheshe yamamodeli e-MoE anezinga lama-trillion ngaphakathi kwesizinda esisodwa somkhawulokudonsa ophezulu, ngaleyo ndlela ithuthukise ukusebenza kahle kokucubungula. #AI大模型##阿里云##月之暗面Kimi#

Ukuhambisana kwe-Day0: Iseva ye-Aliyun Lingjun Zhenwu M890 supernode manje isekela imodeli enkulu i-Kimi K3 ye-Moonshot

I-Aliyun imemezele namuhla ukuthi iseva ye-Lingjun Zhenwu M890 supernode isihambisana ngempumelelo nemodeli enkulu i-Kimi K3 enamapharamitha angu-2.8 trillion. Ngokusebenzisa ukuthuthukiswa okuhlangene kwe-chip, inkundla yokucubungula nemodeli, ukusebenza kahle kokucubungula imodeli kuthuthukisiwe.

Ukuhambisana kwe-Day0: Iseva ye-Aliyun Lingjun Zhenwu M890 supernode manje isekela imodeli enkulu i-Kimi K3 ye-Moonshot

IT之家 ithole ukuthi i-Kimi K3 iyimodeli yakamuva ephambili ye-Moonshot, enobukhulu bamapharamitha obufinyelela ku-2.8 trillion, futhi isebenzisa isakhiwo se-mixture of experts (MoE).

Iseva ye-Lingjun Zhenwu supernode yakhiwe kusetshenziswa i-chip ye-T-Head AI ehlanganisa ukuqeqesha nokusebenzisa, i-Zhenwu M890. Ine-chip yokuxhumana i-ICN Switch 1.0, evumela ama-M890 angu-64 ukuthi afinyelele ku-800 GB/s yokuxhumana okusheshayo kwe-All-to-All. Umthamo wememori yevidiyo ufinyelela ku-9TB, okwenza ithrafikhi ye-EP yokuxhumana kochwepheshe yamamodeli e-MoE anezinga lama-trillion isebenze ngaphakathi kwesizinda esisodwa sokuxhumana esinomkhawulokudonsa ophezulu, kuqinisekiswe ukusebenza kahle kokukhiqizwa kwama-Token.

Ngesikhathi esifanayo, amaqembu e-Aliyun, i-T-Head kanye ne-Kimi asebenzisane ngokujulile emazingeni ama-operator nesitaki sesofthiwe, enza ngempumelelo imodeli ye-Kimi K3 ihambisane kusukela ku-Day0. Isitaki sesofthiwe i-T-Head SAIL se-chip ye-Zhenwu sivumela uhlaka lokucubungula i-Mooncake olwakhiwe yi-Kimi ukuthi lusebenze ngokushesha ngaphandle kokulungiswa okwengeziwe. Ukusekelwa kwe-Triton ku-M890 nakho kusho ukuthi ama-operator amaningi angokwezifiso abhalwe nge-Triton awadingi ukubhalwa kabusha, kunciphise kakhulu umsebenzi wokuhambisana.