ilogo yelamafile
IMozilla yabhengeza, ngeposti yebhlog, ilukukhutshwa komqokeleli, enayo njenge Injongo kukwenza lula ngokumangalisayo ukusetyenziswa kwemizekelo emininzi yeelwimi ezinkulu. (LLM) phantse kuyo nayiphi na idesktop okanye iseva.
Iprojekthi entsha yeMozilla ebizwa ngokuba yi "llamafile", ngumqambi wemithombo evulekileyo engathatha ifayile yeparameter yemodeli yokufunda yomatshini kwifomathi yeGGUF kwaye iyiguqulele kwifayile ephunyezwayo enokuqhuba kwiinkqubo ezintandathu ezahlukeneyo zokusebenza kwi-AMD64 kunye ne-ARM64 hardware.
Kwaye abandivumeli ukuba ndixoke, kodwa bona Iimodeli zolwimi ezinkulu (LLMs) ezisetyenziselwa ukusetyenziswa ekuhlaleni zisasazwa kwiiseti ezahlukeneyo, apho ubunzima bezi fayile zidla ngokuba ziigigabytes ezininzi. Ezi fayile azisetyenziswa ngokwazo ngokuthe ngqo, eyenza nzima ukusasazwa kwayo kunye nokwenziwa xa kuthelekiswa nezinye iintlobo zesoftware. Ukongeza, imodeli ethile inokuthi yenziwe uhlengahlengiso kunye nohlengahlengiso, okukhokelela kwiziphumo ezahlukeneyo xa kusetyenziswa iinguqulelo ezahlukeneyo.
IMozilla ikubonile oku, yathatha inyathelo ngalo mba kwaye yalungisa lo mngeni, Iqela elitsha leMozilla liqalise "llamafile", leyo, njengoko kukhankanyiwe ngasentla, ngumqokeleli oguqula iLLM ibe yifayile yokubini enye ekwaziyo ukusebenza kwiinkqubo ezintandathu ezahlukeneyo zokusebenza (iMacOS, Windows, Linux, FreeBSD, OpenBSD kunye neNetBSD) ngaphandle kwesidingo sofakelo olongezelelweyo. Esi sisombululo siququzelela kakhulu ukuhanjiswa kunye nokusetyenziswa kwe-LLM, ngelixa uqinisekisa ukuhambelana kunye nokuveliswa kwakhona kwenguqulo ethile ye-LLM ngexesha.
Malunga nomqokeleli wellamafile, kukhankanyiwe oko yenziwe ngokudibanisa iiprojekthi ezimbini: llama.cpp (umthombo ovulekileyo wesakhelo sengxoxo ye-LLM) kunye neCosmopolitan Libc (iprojekthi yomthombo ovulekileyo evumela ukuba uqokelele kwaye uqhube iinkqubo zeC kumaqonga amaninzi kunye nezakhiwo). Ngexesha lokuphunyezwa, i-Mozilla ikhankanya ukuba ijongene nemingeni enomdla kwaye kwafuneka yandise kakhulu umda we-Cosmopolitan ukufezekisa iinjongo ezichaziweyo.
Injongo yethu kukwenza imifuziselo emikhulu yolwimi oluvulelekileyo lufikeleleke ngakumbi kubo bobabini abaphuhlisi kunye nabasebenzisi bokugqibela. Senza oku ngokudibanisa i-llama.cpp kunye ne-Cosmopolitan Libc kwisakhelo esidiliza konke ukuntsonkotha kwee-LLM kwifayile enye ephunyezwayo (ebizwa ngokuba yi "llamafile") esebenza kwindawo kwiikhompyuter ezininzi, ngaphandle kofakelo.
Kuyakhankanywa ukuba Enye yeenjongo eziphambili zellamafile yayikukukwazi ukuqhuba kwii-microarchitectures ezininzi ze-CPU. Apha kulapho i-llama.cpp ingena khona ukuze ivumele iinkqubo ze-Intel ezintsha ukuba zisebenzise iimpawu zeprosesa zanamhlanje ngaphandle kokuncama inkxaso kwiikhompyuter ezindala, ngelixa i-AMD64 kunye ne-ARM64 ezi zidibene zisebenzisa i-script yeqokobhe eqala uguqulelo olufanelekileyo. Ifomati yefayile iyahambelana ne WIN32 kunye namaqokobhe amaninzi eUNIX.
Omnye umngeni othe waqwalaselwa ngumba wobunzima beefayile zeLLM., enokuthi idibaniswe kwi-llamafile, ngenxa yenkxaso ye-PKZIP kwilayibrari ye-GGML. Oku kuvumela iintsimbi ezingaxinzelelwanga ukuba zenziwe imephu ngokuthe ngqo kwinkumbulo, njengefayile ezizikhuphayo, kwaye ikwavumela iintsimbi ezibaliweyo ezisasazwa kwi-intanethi ukuba zilungiswe yinguqulelo ehambelanayo yesoftware ebizwa ngokuba yi-cpp, ukuqinisekisa ukuba ukuziphatha okujongwe ekuqaleni kunokuveliswa kwakhona ngokungenasiphelo. .
Masithi unesethi yobunzima beLLM ngendlela yefayile ye-4 GB (kwifomathi yeGGUF esetyenziswa ngokuqhelekileyo). Nge-llamafile ungayiguqula loo fayile ye-4GB ibe yibini esebenza kwiinkqubo ezintandathu zokusebenza ngaphandle kofakelo.
Okokugqibela, kufanelekile ukuba ikhankanywe IMozilla iye yasungula iprojekthi «llamafile» ebhalwe kwi-C/C++ kwaye isasazwe phantsi kwelayisenisi ye-Apache, ethetha izithintelo ezimbalwa ngokweendlela zokusetyenziswa kunye nokufumaneka kwezibonelelo, xa kuthelekiswa neelayisensi ezifana ne-GPL.
Ngokuphathelele unomdla wokufunda ngakumbi ngayo iprojekthi okanye sele ufuna ukuyisebenzisa, ungajongana neenkcukacha kunye/okanye isikhokelo esikhawulezayo osisebenzisayo eli khonkco lilandelayo.