Kwenzekeni
I-llama.cpp ikhiphe inguqulo engu-0.4.0 ku-GitHub ngoSepthemba 4. Ukukhishwa kwengeza usekelo lokuqala lwe-Qwen3.8-Flash-Next, NVIDIA Nemotron-3-Puzzle-75B-A9B, i-DSpark ye-Nemotron 3.5, nanbeige, kanye ne-nanbeige, ne-3B4. DeepSeek-V4-Flash-Vision-Exp. Futhi yengeza amapharamitha wokokufaka kwevidiyo, imikhawulo yokuqukethwe kweseva ye-slot ngayinye, ukufunda i-tensor evilaphayo, izinguquko zendlela yochwepheshe, ukuthuthukiswa kwenqolobane ye-KV, nokulungiselelwa okuningi kwe-backend. Ukukhishwa kubuyekezwa i-ggml isuka ku-0.22.0 iye ku-0.23.0. Le phrojekthi ithi leyo nguqulo yengeza ukunaka okuncane kwe-flash, ukubulawa okuvumelanayo kanye nama-API ancike ekwabeni, umcimbi we-RPC nama-API async, kanye nokusekelwa kwezokuthutha kwe-Apple RDMA. Ikhasi libonisa isakhiwo sasebusuku esihlonzwe njenge-b10809. Ayikubhaleli ukutholakala okupakishwe kanambambili, izimfuneko zokufakwa, ukusatshalaliswa kwesisindo semodeli, noma amanani.
Ikhasi lokukhishwa le-GitHub likhomba i-v0.4.0 njengokukhishwa kwakamuva futhi lirekhoda isikhathi sokushicilelwa sangomhla 4 Septhemba esingu-19:56, ngaphandle kokucacisa izoni yesikhathi embhalweni onikeziwe. Ukukhishwa kufaka phakathi izinguquko ze-API ezifana ne-llama_lazy_mode, izilawuli zosayizi webhafa ye-quantizer, iseshini ebuyekeziwe nezinguqulo zesimo, kanye nabasizi bethokheni ye-multimodal.
Izinguquko zemodeli nezibalulekile zihlanganisa usekelo lokuqala lwe-Qwen3.8-Flash-Next lwezakhiwo, i-NVIDIA Nemotron-3-Puzzle-75B-A9B usekelo, usekelo lwe-DSpark lwe-Nemotron 3.5, usekelo lwe-nanbeige4.2-3B, ukufunda i-tensor evilaphayo, ukubukeka kochwepheshe bokwendlalelo ngalunye, ukuthuthukiswa komlando wochwepheshe be-KV, ukuthuthukiswa komlando wochwepheshe be-KV. ivikela ekuphakameni kwe-RAM ngesikhathi sokulayisha imodeli.
Izinguquko ze-Multimodal neseva zihlanganisa usekelo lwe-DeepSeek-V4-Flash-Vision-Exp, izinketho zomugqa womyalo wevidiyo, imikhawulo yokuqukethwe kwendawo ngayinye, ama-URL edatha yemidiya, ukulondolozwa okuzenzakalelayo kokuphumayo kokucabanga, kanye nokwenqatshwa kwezingcingo zamathuluzi ezisiza ezigcwaliswe kusengaphambili. I-UI iphinde iguqule inqubomgomo yamathuluzi nokuziphatha kwezilungiselelo, kodwa umthombo awuhlinzeki ngedatha yokutholwa komsebenzisi noma yokusebenza.
Imininingwane yomthombo: github.com ↗
Kungani kubalulekile
Lesi isibuyekezo esikhulu sesikhathi sokusebenza somthombo ovulekile esisetshenziswa onjiniyela abakha i-AI kanye nezinhlelo zokusebenza ze-multimodal. Ukufakwa kwemodeli yayo enwetshiwe kungenza amamodeli asanda kukhishwa ahlolwe ngaphakathi kwezinhlelo ezisekelwe ku-llama.cpp, kuyilapho ukusekelwa kokokufaka kwevidiyo kunweba izinhlobo zemidiya lezo zinhlelo ezingase zicushwe. Umthombo uchaza ukusebenza kanye nomsebenzi ohlobene nenkumbulo, kodwa awunikezi amabhentshimakhi azimele, ngakho izinzuzo ezingokoqobo zizoncika kuhadiwe, amafomethi wamamodeli, nokucushwa kokuphakwa.
Ukukhishwa kuxhumanisa amamodeli ezakhiwo amaningana amasha ku-codebase esetshenziswa kakhulu, efaka usekelo lokuqala lwe-Qwen3.8-Flash-Next ne-Nemotron-3-Puzzle. Lokho kungase kunciphise umsebenzi wokuhlanganiswa konjiniyela abazama lawo mamodeli, nakuba umthombo ungaqalisi ukuhambisana kuyo yonke inkundla noma ukulungiselelwa.
Isibuyekezo se-ggml sengeza ingqalasizinda yokunakwa okungatheni, ama-backends asynchronous, imicimbi yekholi yenqubo ekude, kanye ne-Apple RDMA. Lezi zinguquko zingaba nendaba ekusetshenzisweni okungasebenzisa lezo zingemuva ezithile, kodwa ikhasi lokukhishwa alibali ukubambezeleka, ukuphuma, ukusetshenziswa kwememori, noma ukuthuthukiswa kokwethembeka.
Amapharamitha wevidiyo, ukusekelwa kwedatha-URL yemidiya, kanye nezinguquko zokucutshungulwa kwangaphambili kwe-multimodal kunikeza indlela ebambekayo yezinhlelo zokusebenza eziphethe ividiyo neminye imidiya. Umthombo awusho ukuthi lawa makhono ayatholakala ekwakhiweni okupakishiwe noma ngaphansi kwayiphi imikhawulo yemodeli ethile.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ukukhishwa kokulandelela, ukuhlolwa okuqondile kwe-backend, kanye nemibhalo kufanele kucacise ukuthi imodeli entsha nezici zevidiyo ziziphatha kanjani kuwo wonke amahadiwe asekelwayo. Ukungaqiniseki okushesha kakhulu ukuthi ukuqaliswa kokuqala kwe-Qwen3.8-Flash-Okulandelayo kuyawuthola yini umsebenzi othenjisiwe wokuthuthukisa nokuthi ukunaka okuncane nezinguquko zenkumbulo zithuthukisa kangakanani umthwalo wangempela wokusebenza.
Buka ukuze uthole izibuyekezo zokuthuthukisa ku-Qwen3.8-Flash-Next kanye nezilungiso ezengeziwe zezindlela ezintsha zemodeli ye-multimodal.
Buka imiphumela yebhentshimakhi ehlukanisa izimangalo zokusetshenziswa kwekhasi lokukhishwa ezinzuzweni ezilinganiselwe ngesivinini, ukusetshenziswa kwememori, nokuvumelana.
Buka amadokhumenti okufinyelela okupakishiwe, amafayela emodeli asekelwayo, izimfuneko zezingxenyekazi zekhompuyutha, kanye nokulungela ukukhiqiza. Leyo mininingwane ayinikeziwe yikhasi lokukhishwa.