Kungani ama-LLM ombhalo emakhulu kangaka?
Ukugijima gpt-6b idinga i-12GB yenqama ngenkathi gpt-20b futhi na ngaphezulu.
Uma kuqhathaniswa ukusabalalisa okuzinzile edinga kusuka ku-3GB kuya ku-5GB.
Yini enomthelela ezidingweni zenkumbulo ukuze usebenzise imodeli? Inombolo yamapharamitha, njengoba yonke ipharamitha iyi-float (16/32/64 bits).
Isithombe sibiza izinkulungwane zamagama? Ingabe Izithombe ziyithuluzi elingcono lokucindezela ulwazi kunezilimi? Kungenzeka. Futhi mhlawumbe izithombe azicacile ngokujwayelekile kunombhalo.
Cabanga ukuthi isethi yokuphumayo ilungisiwe (Indawo yonke). Singachaza isithombe ngemishwana ehlukene futhi yonke ingaba okokufaka okwamukelekayo, noma singamukela izithombe eziningi ezisuka ekufakweni okufanayo kombhalo njengezivumelekile.
Nakuba ukuthola okuphumayo kombhalo okwamukelekayo kokunye okokufaka kombhalo kufanele kulandele incazelo yokufaka kwawo ngemithetho eqinile. Lokhu kusho ukuthi okokufaka okumbalwa kwemephu kusethi egcwele yokuphumayo.
Ngamanye amagama, i-LLM yokuguqula umbhalo uye esithombeni inokuphezulu ukuzimela kunombhalo-ube-umbhalo we-LLM ngenxa yalokho okuhlukanisa njengokuphumayo okwamukelekayo kuncike emikhawulweni yethu njengezihlukanisi zezithombe "zezikhwama zenyama". Sicubungula isithombe esihle esiphuma kumbhalo, uma leso sithombe siphethe incazelo ebanzi yokwaziswa kwethu, sishiya okuningi "ekuhumusheni". Ngenkathi sifunda umbhalo, silindele ukuthi ulandele imithetho eqinile yohlelo, futhi ube nencazelo.
Singakwazi ukucacisa ukungapheleli ezithombeni ngoba akubona bonke abantu abakwazi ukudweba futhi adale izithombe, kuyilapho wonke umuntu engakwazi ukufunda nokubhala iphrozi ehloniphekile. Ukusetshenziswa kwama-LLM kubonakala emandleni ethu okufeza umsebenzi i-LLM yakhelwe ukuwufeza. Uma singakwazi ukwenza okuthile, ama-LLM kufanele abe sezingeni eliphansi ukuze abe usizo, amapharamitha ambalwa azowadinga. Ngamanye amazwi singabadwebi be-crap! futhi ibha yethu yemodeli "enhle" yombhalo uye esithombeni isethwe ngaphansi. Ukube wonke umuntu ubengumhlaziyi/umkhiqizi wesithombe okahle kangangokuthi izidingo zethu zesithombe esihle beziqinile, amamodeli ombhalo-ukuya-isithombe bekuyodingeka abe makhulu kakhulu kunokuguqula umbhalo ube umbhalo, ngenxa yokuthi kungakanani ukuhlaziya okudingekayo ukuze sihlukanise ngempela isithombe kuyo yonke incazelo yaso.
Yebo, awekho ama-LLM okuthumela isithombe noma umbhalo-kuya-umbhalo anencazelo yangempela, kodwa kulula ukukhohlisa incazelo ngezithombe kunokuba kunjalo ngombhalo ngoba izithombe zingamazinga aphezulu okukhipha kunombhalo. Umbhalo uwuhlobo lokuxhumana oluwuchungechunge lapho ulwazi olusha ludluliselwa khona ngokuhlanganisa ukulandelana kwamaqoqo amathokheni anencazelo ethile (zibize ngezigaba).
Izithombe esikhundleni salokho zinezakhiwo ze-fractal, incazelo ayincikile kwelinye iqoqo elaziwayo endaweni ethile yangaphambilini yokulandelana, kodwa kanyekanye ukuma kwenqwaba yamaqoqo ngaphakathi kwekhanvasi.
Ukuze "uqonde" isithombe kufanele udale amaqoqo, ngaphandle kokwazi ukuhleleka okuthile, ulandele kuphela ukwehlukana kwemibala, futhi uhlukanise lawo maqoqo (ngamabhizinisi, umumo, ukujula, umbani, njll), yenza kanjalo ezingeni ngalinye lokuqanjwa (kusuka ku-1) amaphikseli endaweni egcwele yekhanvasi), bese unikeza izisindo kukho konke ubudlelwano obukhona phakathi kwawo wonke amazinga okuhlanganisa akhiqizwayo. Lezi zidinga okungenani amamodeli amakhulu angu-3, kanye no-1 ngaphezulu uma ufuna ukwenza inqubo yokuhlukanisa ngokukhethekileyo.
Lokhu akukhona okwenziwa ukusakazeka okuzinzile...kancane nje, “Inembiza” yezisindo (isikhala esicashile) esibamba ulwazi olubhalwe ngekhodi (olunomsindo) lwazo zonke izithombe eqeqeshwa ngalo. Ngingasho ukuthi imodeli isebenzisa izici zezibalo zemibala ukubikezela isithombe esivela kumbhalo (noma esinye isithombe). Ngamanye amazwi, iphendula umbuzo othi:
Isiphi isithombe okungenzeka kakhulu esenelisa ukuphindaphinda kokususa umsindo, kusukela ekwazisweni okunikeziwe?
Leli iqhinga elihlakaniphile lokucindezela ulwazi oluningi futhi yini evumela ukusakazeka okuzinzile ukuthi kulingane ku-5GB wenqama, kuyilapho indlela engichaze ngayo ekuqaleni ingadinga okungenani izikhathi ezi-3 lelo nani (kodwa mhlawumbe okuningi kakhulu ekusebenzeni).
Kuyadabukisa ukuthi umthetho ka-moore (noma i-litography, uma ukhuluma ngamanani ezinqama) wakhuphuka kahle lapho ukuthuthukiswa okungengaphezu kuka-5x bekuzovumela ukusebenzisa ama-LLM asebenza kahle kakhulu endaweni kuwo wonke umuntu wemishini!
Ngokuthuthuka okumbalwa kwezingxenyekazi zekhompuyutha zethu kungase kuhlakaniphe kakhulu ngokushesha okukhulu.
