Kwenzekeni
Iphrojekthi ye-llama.cpp ikhiphe inguqulo engu-0.3.0 ngo-Aug. 25, ngokusho kwekhasi layo lokukhishwa le-GitHub. Isibuyekezo sengeza ukusekelwa kwemodeli ye-multimodal yama-dots3-note, okuhlanganisa uhlobo olusha lwenqolobane yokhiye we-DSA-ISWA, futhi sengeza umbono nokubamba umsindo kuleyo modeli. Iphinde yethule imodi yokuhlukanisa i-tensor ye-DeepSeek 4, ilungise ukubuyisela emuva kokulandelana okuningi, futhi yengeze ukwesekwa kokuqagela kwamathokheni amaningi kwe-GLM-4.5-Air.
Ikhasi le-GitHub likhomba i-v0.3.0 njengokukhishwa kwakamuva kwekhosombe lomphakathi le-llama.cpp futhi lirekhoda ukukhishwa ngo-10:22 ngo-Aug. 25. Ukukhishwa kwethulwa njengesibuyekezo senguqulo enkulu kunesiqephu esincane sokulungisa. Izinguquko zayo ezimaphakathi zithinta ukuthi isofthiwe iwasekela futhi iwasebenzise kanjani amamodeli e-AI, okuhlanganisa imodeli entsha ye-multimodal, izici zokubhekisela eziqondene nemodeli ethile, i-tensor parallelism, nezinguquko ezingxenyeni ezabiwe ezisetshenziswa iseva namanye amathuluzi. Isengezo semodeli evelele kakhulu yi-dots3-note. Amanothi okukhishwa athi i-llama.cpp yengeza imodeli kanye nohlobo olusha lwenqolobane yokhiye we-DSA-ISWA. Bafaka uhlu ngokuhlukene ukusekelwa kombono nomsindo wenothi lama-dots3, izithombe ze-WebP nge-ffmpeg, kanye ne-algorithm yokushintsha usayizi enembile yomcamelo. Amanothi aphinde athi ukulayishwa kwevidiyo kwalungiswa lapho i-athomu ye-moov yefayela ivela ekugcineni. Lezi zinguquko zibonisa ukuthi iphrojekthi inwebela ngale kokusetshenziswa kwemodeli egxile embhalweni iye kusethi ebanzi yokokufaka kwemidiya, nakuba umthombo ungachazi amakhono emodeli noma izinhlelo zokusebenza ezihlosiwe.
Ku-DeepSeek 4, inguqulo 0.3.0 yengeza imodi yokuhlukanisa i-tensor ngenketho ethi “-sm tensor” futhi ilungisa ukubuyisela emuva lapho ukulandelana okuningi kusebenza. Ukukhishwa futhi kubika ukulungiswa kokusabalaliswa kwe-tensor-split state ngokusabalaliswa kwe-tensor-parallel. Ngokwehlukana, i-GLM-4.5-Air ithola ukwesekwa kokubikezela kwamathokheni amaningi, kuyilapho i-bailingmoe3 ithola ukwesekwa kwe-DSpark. Lezi izinguquko zezinga lokuqaliswa ezithinta indlela amamodeli athile angenziwa ngayo, afaniswane, noma asheshiswe; ukukhishwa akusho ukuthi kuthuthukisa ikhwalithi yemodeli.
Isibuyekezo siphinda sishaye ingxenye ye-ggml ibe yi-v0.22.0. Amanothi achaza ukwesekwa kwe-tensor-split ku-backend ye-multi-backend meta ye-ggml, ukusabalalisa kwe-split-state okuthuthukisiwe, ukusebenza ngakunye Imithombo yensimbi enokuhlanganiswa okufanayo, kanye nokulungiswa okwenza i-ggml_clamp isebenze okungekho endaweni. Izinguquko ezengeziwe zihlanganisa imisebenzi emisha, i-Q2_K SYCL kernels, inhlanganisela yochwepheshe be-OpenCL exubile, nokulungiswa kuyo yonke i-CUDA, Metal, SYCL, Vulkan, OpenCL, neWebGPU. Iseva izuza okuguquguqukayo kwendawo ukuze kunwebe umehluko wokulungisa iphutha, futhi isixhumi esibonakalayo sewebhu sizuza ukuzulazula kwengxoxo ngamathebhu.
Imininingwane yomthombo: github.com ↗
Kungani kubalulekile
Ukukhishwa kunweba ububanzi bamamodeli we-AI nezinhlobo zokufakwayo ezisekelwa ukusetshenziswa komthombo ovulekile osetshenziswa kabanzi kuyilapho kubhekwana nenkumbulo, ukufana, nezinkinga ezingemuva ezihilelekile ekusebenziseni amamodeli amakhulu noma ayinkimbinkimbi kakhulu. Umthelela osebenzayo uzoncika kuhadiwe, amafayela oyimodeli, ukucushwa kokwakha, kanye ne-backend ethize esetshenzisiwe.
Ukubaluleka okusebenzayo ukuthi isitaki sokusebenza semodeli yomthombo ovulekile sengeza usekelo lwezakhiwo zamamodeli ahlukahlukene kanye nezindlela ekukhishweni okufanayo. Konjiniyela nabacwaningi abasebenzisa i-llama.cpp, usekelo lombono we-dots3-note nokokufaka komsindo kunganciphisa isidingo sokugcina izindlela ezihlukene zokusayinda zalokho okokufaka. Amanothi okukhishwa asungula ubukhona bosekelo, kodwa awaqinisekisi ukuthi luphelele kangakanani, lushesha kangakanani, noma luthembekile kangakanani kuzingxenyekazi zekhompyutha ezithile.
Ukuhlukaniswa kwe-tensor kubalulekile ngoba kuthinta ukuthi umthwalo wemodeli noma isimo usatshalaliswa kanjani kuwo wonke amadivayisi. Izinguquko ze-DeepSeek 4 zingenza isofthiwe isebenziseke kakhudlwana kubantu abakhompuyutha yabo etholakalayo ihlukaniswa kumadivayisi amaningi, kuyilapho ukulungiswa kokuhlehlisa kungase kubaluleke emithwalweni yomsebenzi egcina ukulandelana okungaphezulu kokukodwa okusebenzayo. Lezo izinzuzo zokusebenza ezibambekayo ezisekelwe ezinguqukweni ezisohlwini, hhayi emiphumeleni elinganiselwe: umthombo awunikezi umphumela, inkumbulo, ukubambezeleka, noma izinga lokuhluleka ukuqhathanisa nenguqulo 0.2.0.
Izinguquko ze-ggml zinweba ukufinyelela kokukhishwa kuzo zonke izingxenye ezingemuva zehadiwe. Ukuhlanganiswa okuhambisanayo kwemithombo Yensimbi kungase kuthinte izikhathi zokwakha, kuyilapho umsebenzi wokuhlukaniswa kwe-meta-backend nokulungiswa kwe-CUDA, SYCL, Vulkan, OpenCL, kanye ne-WebGPU zikhuluma ngesendlalelo sokuphatheka esixhumanisa amamodeli nezinhlobo ezahlukene zehadiwe. Ukukhishwa futhi kuklelisa inguquko ye-Mamba2 ehloselwe ukuthumela i-GEMM esikhundleni se-GEMV, kodwa ayibali noma iyiphi inzuzo yokusebenza. Ngakho-ke abasebenzisi kufanele baphathe amanothi njengoshintsho lokusebenzisa, hhayi njengombiko wokulinganiswa.
Ukukhishwa kubalulekile ku-AI ecosystem ebanzi ngoba ukusekelwa kwemodeli akulolongwa abadali bamamodeli kuphela kodwa futhi kusendlalelo sesofthiwe esenza amamodeli asebenziseke kumasistimu ahlukene. Ukwengeza imodeli noma indlela ethile kungaba nomthelela wokuthi imaphi amaphrojekthi asebenzayo ekusetshenzisweni kwendawo noma okukhethekile. Noma kunjalo, okubalulekile okungaziwa kusasele: umthombo awusho imigomo yelayisensi noma yokusabalalisa yama-dots3-nothi, amamodeli esisindo adingekayo, amasistimu okusebenza asekelwayo, ubuncane bezingxenyekazi zekhompuyutha, amafomethi asekelwayo omsindo nevidiyo ngale kokulungiswa okusohlwini, noma ukuthi zonke izici ziyatholakala yini kumpahla yasebusuku.
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
Amanothi okukhishwa awanikezi amabhentshimakhi azimele, izimfuneko zehadiwe, ama-matrics ahambisanayo, noma ubufakazi bokukhishwa kokukhiqiza. Ukulandelela okubalulekile ukuthi ingabe abasebenzisi babika ukusebenza okuthembekile kombono nezindlela zomsindo we-dots3-note, i-DeepSeek 4 ihlukanisa i-tensor kuwo wonke amadivayisi, kanye ne-CUDA, Metal, SYCL, Vulkan, OpenCL, ne-WebGPU ebuyekeziwe.
Iphuzu lokuqala lokuqinisekisa ukumbozwa okusebenzayo kwe-dots3-note. Amanothi okukhishwa kwegama lokubona nokusekelwa komsindo, ukuqoshwa kwe-WebP, ukulungiswa kokulayisha ividiyo, nokuziphatha okushintshile, kodwa akubandakanyi izibonelo zokuhlola noma amathebula okokufaka asekelwe. Ukuhlola okuzimele kufanele kuqinisekise ukuthi imaphi amafomethi asebenza, ukuthi ukucubungula kusengaphambili kuyithinta kanjani imiphumela, ukuthi umsindo nokubona kungahlanganiswa yini, nokuthi ingabe ukuziphatha kuyahambisana yini kuzo zonke izingemuva ezisekelwayo. Kuze kube yileso sikhathi, “ukwesekwa” kufanele kuqondwe njengesimangalo sokuqaliswa kwezinga lephrojekthi kunobufakazi bokulungela ukukhiqiza.
Udaba lwesibili umphumela womhlaba wangempela wokuhlukaniswa kwe-tensor ye-DeepSeek 4. Abasebenzisi bazodinga ukunquma ukuthi ingabe i-“-sm tensor” isebenza kuwo wonke amadivayisi abo, ukuthi inkumbulo ihlukaniswa kanjani, kanye nokuthi ukubuyisela emuva kokulandelana okuningi kuhlala kuthembekile yini ngaphansi komthwalo omude noma ohambisanayo. Umthombo awunikezi uhlu lwezingxenyekazi zekhompuyutha, idatha yokusebenza, noma ukuhlaziya amaphutha. Imibiko evela kubanakekeli nabasebenzisi izoba usizo ikakhulukazi ngoba ukuziphatha kwe-tensor-parallel kungahluka ngezishayeli, izici ezihlanganisiwe, inhlanganisela yedivayisi, nokucushwa kwemodeli.
Indawo yesithathu i-backend parity. Amagama okukhishwa abuyekeza i-CUDA, Metal, SYCL, Vulkan, OpenCL, neWebGPU, kodwa ayisho ukuthi yonke imodeli entsha noma ukusebenza kusebenza ngokulinganayo kungemuva ngalinye. I-changelog ngokwayo iphawula ukuthi ukuhlolwa kwe-dots3-note architecture kukhutshaziwe ku-WebGPU, okuyisizathu esiphathekayo sokugwema ukuthatha ukufakwa okuphelele kwe-cross-backend. Amadokhumenti okulandelela noma imiphumela yokuhlolwa kufanele icacise ukuthi yini enikwe amandla, ehlolwayo, ekhawulelwe, noma engatholakali.
Okokugcina, ukukhishwa okulandelayo kwephrojekthi kuzobonisa ukuthi lezi zinguquko zihlala zizinzile yini. Izinto okufanele ziqashwe zihlanganisa ukulungiswa kokulayishwa kwenothi lama-dots3 kanye nokucubungula kusengaphambili kwemidiya, ukuhlehla okuhlanganisa ukubikezela kwamathokheni amaningi noma ukushumeka, izinguquko eziqhubekayo ze-DeepSeek 4 parallelism, kanye nezibuyekezo kumqondo wokungena kwesikhala seseva. Ukukhishwa kufaka phakathi impahla yokwakha yasebusuku, kodwa umthombo awuchazi ukupakishwa kwayo, ukukhiqizwa kabusha, noma ubudlelwano nezakhiwo ezizinzile. Abukho ubufakazi obuzimele ezintweni ezinikeziwe obusungula ukutholwa, isikali sokuphakelwa, noma umthelela wabasebenzisi.