IMoore Threads ikhiphe ngokusemthethweni i-MTT S5000 Prefill-as-a-Service Technical White Paper, yamemezela uhlobo olusha lwe-“Prefill As a Service” olwakhiwe ngekhadi lekhompyutha elihlakaniphile le-AI MTT S5000 elihlanganisa ukuqeqesha nokusebenzisa amamodeli, olubhekiswe ezimweni zokucabanga ngomongo omude ezifana nama-AI Agent, ukukhiqizwa kwekhodi nokuhlaziywa kwemibhalo emide kakhulu, lunikeze imboni indlela esebenzayo ehlanganisa ukusebenza kahle kokucabanga, ukulawulwa kwezindleko kanye nokwenza imali ngamandla ekhompyutha.
Ngokwesethulo seMoore Threads, njengoba usayizi womongo wokufaka wamamodeli amakhulu ukhula ngokushesha ufinyelela kumakhulu ama-K kuya ezingeni le-1M, ukusetshenziswa kwamandla ekhompyutha kanye nengcindezi yokubambezeleka kohlamvu lokuqala okubangelwa ukufakwa kwemibhalo emide kuyaqhubeka nokukhula. Lokhu sekuya kuba yisithiyo esibalulekile esivimba ukusebenza kahle kwezinsizakalo zokucabanga zamabhizinisi kanye neTotal Cost of Ownership (TCO). Indlela yendabuko yokusebenzisa amandla ekhompyutha kumaqoqo ehardware afanayo ayisakwazi ukubhekana nezidingo ezahlukene zezinsiza zeh hardware ezigabeni ezehlukene zokubala.
Inkinga esemqoka iwukungahambisani kwesakhiwo: ekucabangeni okuku-inthanethi kwamamodeli amakhulu, isigaba sokufaka i-Prefill (ukugcwalisa kusengaphambili) siwumsebenzi odinga kakhulu amandla okubala, futhi siboshwe amandla okubala izinombolo zamaphuzu antantayo; isigaba sokukhipha i-Decode (ukuhlukanisa nokukhipha) siwumsebenzi odinga kakhulu ukufinyelela enkumbulweni, futhi siboshwe umkhawulokudonsa wememori yevidiyo.
Ukusebenzisa kokubili ndawonye echibini elilodwa lezinsiza zomzimba kuholela ekumosheni okubili: uma ihardware ikhethwa ngokomkhawulokudonsa we-Decode, imemori enkulu ye-Prefill isebenza ngaphansi kwamandla isikhathi eside; uma ikhethwa ngokwamandla okubala e-Prefill, amayunithi okubala e-Decode ahlala engasebenzi kakhulu. I-Technical White Paper iveza ukuthi ukhiye wokunqoba lesi sithiyo uwukuhlukanisa ngokuphelele i-Prefill ne-Decode, ukuze ngayinye isebenze echibini lezinsiza zehardware elihambisana kangcono nezici zayo zokubala:
Ichibi lamandla okubala le-Prefill: ligxile ekusetshenzisweni okuphezulu kwamandla okubala kanye nokubambezeleka okuphansi kakhulu kohlamvu lokuqala (TTFT);
Ichibi lezinsiza le-Decode: ligxile ekusebenzeni okuhambisana nabasebenzisi abaningi kanye nokubambezeleka okuzinzile kwe-Token ngayinye (ITL).
Ngokutshala imali okuhlukaniswe ngokwezigaba zehardware, ukumiswa kwamandla okubala kushintsha kusukela “ekusetshenzisweni okweqile okuvamile” kuya “ekufaneni nezidingo”, ngaleyo ndlela kwehlise izindleko zengqalasizinda nge-Token ngayinye, kanti kugcinwa imingcele ye-Service Level Objective (SLO).