Chii chaitika
llama.cpp yakabudiswa vhezheni 0.4.0 paGitHub musi waGunyana 4. Kuburitswa kunowedzera rutsigiro rwekutanga rweQwen3.8-Flash-Next, NVIDIA Nemotron-3-Puzzle-75B-A9B, DSpark yeNemotron 3.5, nanbeige, uye nanbeige, uye. DeepSeek-V4-Flash-Vision-Exp. Iyo inowedzerawo vhidhiyo-yekupinda ma paramita, per-slot server mamiriro emiganhu, usimbe kuverenga, nyanzvi-nzira shanduko, KV-cache kuvandudzwa, uye akawanda backend optimizations. Iyo yekuburitsa inogadziridza ggml kubva 0.22.0 kusvika 0.23.0. Iyo purojekiti inoti iyo vhezheni inowedzera sparse flash kutarisisa, asynchronous kuuraya uye kugovera-kutsamira APIs, RPC chiitiko uye async APIs, uye Apple RDMA yekufambisa rutsigiro. Iro peji rinonyora chivakwa chehusiku chinozivikanwa se b10809. Iyo haina magwaro epakeji-binary kuwanikwa, kuisirwa zvinodiwa, modhi-huremu kugovera, kana mitengo.
GitHub peji rekuburitsa rinoratidza v0.4.0 seyachangoburwa uye inorekodha nguva yekuburitswa yaGunyana 4 nguva dza19:56, pasina kudoma nguva yenguva muzvinyorwa zvinopihwa. Kuburitswa kunosanganisira API shanduko dzakadai se llama_lazy_mode, quantizer buffer-saizi zvidzoreso, yakagadziridzwa chikamu uye nyika shanduro, uye itsva multimodal tokenization vabatsiri.
Kuchinja kwemuenzaniso nepakati kunosanganisira yekutanga Qwen3.8-Flash-Next architecture rutsigiro, NVIDIA Nemotron-3-Puzzle-75B-A9B tsigiro, DSpark tsigiro yeNemotron 3.5, nanbeige4.2-3B tsigiro, simbe yekuverenga, per-layer-kutarisisa kwenyanzvi, kuvandudza nhoroondo yeVV, kugadziridza kwehunyanzvi hwekuita, nstov inodzivirira kubva pakakwirira RAM panguva yekurodha modhi.
Multimodal uye server shanduko dzinosanganisira DeepSeek-V4-Flash-Vision-Exp tsigiro, vhidhiyo yekuraira-mutsara sarudzo, per-slot mamiriro emiganhu, data maURL enhau, kuchengetedza kwakasarudzika kwekubuda kwekufunga, uye kurambwa kwekufanozadzikiswa kwechishandiso chekubatsira mafoni. Iyo UI zvakare inoshandura-chishandiso-gwara uye marongero maitiro, asi sosi inopa hapana kutorwa kwemushandisi kana data rekuita.
Kwakabva mashoko: github.com ↗
Nei zvichikosha
Ichi chiitiko chakakura kune yakavhurika-sosi yekumhanyisa nguva inoshandiswa nevagadziri vanovaka AI inference uye multimodal application. Yakawedzerwa modhi yekuvhara inogona kuita kuti ichangoburwa mamodheru iongororwe mukati mellama.cpp-based masisitimu, nepo vhidhiyo-yekupinza rutsigiro ichiwedzera mhando dzemidhiya idzo masisitimu anogona kugadzirisa. Sosi yacho inotsanangura kuita uye basa rine chekuita nendangariro, asi haripe mabhenji akazvimiririra, saka budiriro inoshanda inotsamira pane Hardware, mafomati emhando, uye deployment kumisikidza.
Kuburitswa kunobatanidza akati wandei mavakirwo emhando nyowani kune yakashandiswa zvakanyanya inference codebase, kusanganisira yekutanga Qwen3.8-Flash-Inotevera uye Nemotron-3-Puzzle rutsigiro. Izvo zvinogona kuderedza basa rekubatanidza kune vanogadzira vanoedza neaya mamodheru, kunyangwe sosi yacho isingagadziri kuenderana papuratifomu yega yega kana kumisikidzwa.
Iyo ggml yekuvandudza inowedzera zvivakwa zvekutarisisa zvishoma, asynchronous backends, kure maitiro ekufona zviitiko, uye Apple RDMA. Shanduko idzi dzinogona kuve nebasa kune deployments inogona kushandisa iwo chaiwo backends, asi peji rekuburitsa hariverenge latency, throughput, ndangariro kushandiswa, kana kuvimbika kuvandudzwa.
Vhidhiyo paramita, data-URL rutsigiro rwemidhiya, uye multimodal preprocessing shanduko inopa yakawedzera kongiri nzira yezvishandiso zvinobata vhidhiyo uye mamwe midhiya. Iyo sosi haitaure kana aya masimba anowanikwa mumapakiti ekuvaka kana pasi pezvipimo-zvakasarudzika.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Kuteedzera-kuburitswa, backend-chaiyo bvunzo, uye zvinyorwa zvinofanirwa kujekesa maitiro matsva emhando uye vhidhiyo maficha pane ese anotsigirwa hardware. Kunyanya kukurumidza kusava nechokwadi ndechekuti yekutanga Qwen3.8-Flash-Inotevera kuisirwa inogamuchira yakavimbiswa optimization basa uye kuti yakawanda sei iyo sparse-kutarisisa uye ndangariro shanduko inovandudza basa chairo.
Tarisa kune optimization updates kuQwen3.8-Flash-Inotevera uye zvimwe zvigadziriso zvezvichangobva kuwedzerwa multimodal modhi nzira.
Tarisa uone mabhenji mabhenji anoparadzanisa kuburitswa kwepeji yekuburitsa zvichemo kubva kuyerwa kuwana mukumhanya, kushandiswa kwendangariro, uye concurrency.
Tarisa zvinyorwa pamusoro pekupinda mupakiti, anotsigirwa modhi mafaera, hardware zvinodiwa, uye kugadzira-kugadzirira. Iwo mameseji haana kupihwa neiyo yekuburitsa peji.