Cishe usuvele ujwayelene namathuluzi afana ne-ChatGPT, i-Claude, noma i-Gemini. Nakuba emangalisa, kukhona okubambekayo: asebenza efwini. Lokhu kusho ukuthi wonke amagama owathayiphayo aya kumaseva enkampani, okungaba yingozi enkulu yobumfihlo uma uphatha idatha ebucayi noma usebenza emikhakheni lapho umthetho uvimbela khona ulwazi ukuthi luphume ehhovisi.
Yilapho u-Ollama engena khona, uhlelo lokusebenza oluguqula ikhompyutha yakho ibe yisikhungo sezinzwa se-AI . Khohlwa ukubhalisa kwanyanga zonke futhi uthembele ekuxhumekeni kwe-inthanethi okuzinzile; ngaleli thuluzi, ungalanda amamodeli omthombo ovulekile bese uwasebenzisa ngqo kwihadiwe yakho, ugcine ukulawula okuphelele kwedatha yakho.
Iyini ngempela i-Ollama futhi yiziphi izinzuzo zayo?
I-Ollama iyisofthiwe yomthombo ovulekile esebenza njengeklayenti lokuphatha amamodeli ezilimi ezinkulu (i-LLM). Inzuzo yayo eyinhloko ukuthi yenza kube lula ukuguquguquka kobuchwepheshe , ukuphatha ukulungiswa kwe-GPU kanye nokuphathwa kwememori ngakho udinga nje ukwenza umyalo bese uqala ukuxoxa.
Izinzuzo zicacile. Okokuqala, ubumfihlo buphelele njengoba kungekho lutho oluphuma emshinini wakho. Okwesibili, wonga ezindlekweni zethokheni ye-API noma ezindleleni zanyanga zonke zezinhlelo ze-Plus. Okwesithathu, uma ithempulethi isiku-hard drive yakho, ungayisebenzisa ungaxhunyiwe ku-inthanethi ngokuphelele , ilungele uma usendizeni noma ezindaweni ezingenayo inethiwekhi eningi.
Kodwa-ke, akusikho konke ukukhanya kwelanga kanye nothingo. Inkinga enkulu ukuthi udinga ihadiwe enamandla . Uma uzama ukusebenzisa imodeli enkulu kwi-PC ene-RAM encane, impendulo izoba kancane kangangokuthi uzoba nesikhathi sokwenza ikhofi ngaphambi kokuba iqede umusho wayo. Ngaphezu kwalokho, i-AI yazi kuphela lokho ekufundile kuze kube usuku lwayo lokuqeqeshwa, ngakho ayinakho ukufinyelela ezindabeni eziphuthumayo ngesikhathi sangempela.
Izidingo zesistimu kanye nehadiwe enconywayo
Ukuze ugweme okuhlangenwe nakho okukhungathekisayo, kufanele ubheke izingxenye zakho. Umthombo obaluleke kakhulu yi -RAM kanye ne-VRAM yekhadi lakho lezithombe . Njengomthetho ojwayelekile, imodeli engu-1.5B ithatha cishe isikhala esingu-1GB, kodwa idinga imemori eyengeziwe ukuze isebenze kahle.
- Amamodeli Amancane (270M kuya ku-4B): Kulungele imishini ephansi noma ephathwayo.
- Amamodeli Amancane (4B kuya ku-14B): Inketho elinganiselayo yomsebenzisi wasekhaya.
- Amamodeli Aphakathi (14B kuya ku-70B): Lapha udinga ihadiwe engathi sína.
- Amamodeli Amakhulu (angaphezu kuka-70B): Kuphela kwemishini enezinsizakusebenza ezinkulu.
Ngokuphathelene ne-GPU, ukuba nekhadi le-NVIDIA eline-CUDA, ikhadi le-AMD eline-ROCm, noma i-Mac ene-Apple Metal kwenza umehluko omkhulu, kusheshisa ukwenziwa kombhalo izikhathi ezi-5 kuya kweziyi-10 uma kuqhathaniswa nokusebenzisa i-CPU kuphela. Uma uzothenga imishini, funa amakhadi ehluzo ane-VRAM okungenani engu-16GB ukuze ungaphelelwa yisikhathi ngokushesha.
Umhlahlandlela wokufaka wesinyathelo ngesinyathelo
Ukufaka i-Ollama kulula ngokumangazayo futhi kungenziwa ngemizuzu embalwa kuye ngohlelo lokusebenza olusebenzisayo:
Ku -Windows , mane nje ulande ifayela le-.exe kuwebhusayithi esemthethweni. Ngokuvamile lizofakwa kufolda ye-AppData yomsebenzisi. Kulabo abathanda izitsha, lingaphinde lisetshenziswe kusetshenziswa i-Docker Desktop ngokusebenzisa umyalo wokufaka osemthethweni ku-terminal.
Okwabasebenzisi be LinuxIndlela esheshayo ukusebenzisa i-terminal enomyalo curl -fsSL https://ollama.com/install.sh | shUma isifakiwe, isevisi ivame ukusebenza ngemuva. Uma udinga ukushintsha lapho amamodeli agcinwa khona ukuze ugweme ukugcwalisa idiski yakho eyinhloko, ungahlela i-environment variable. OVEN_MODELS kufayela lokucushwa kwesevisi yesistimu.
En macOSUkufakwa kulula kakhulu kusetshenziswa isifaki esilandiwe noma ngisho nokusebenzisa brew install ollamaUhlelo luzosebenzisa ngokuzenzakalelayo i- Ukusheshisa kwensimbi ama-chips e-Apple Silicon.
Indlela yokuphatha nokusebenzisa amamodeli e-AI
Uma iseva ye-Ollama isisebenza (uzoyibona kusithonjana sebha yomsebenzi), ungaqala ukulanda amamodeli. Umyalo oyisisekelo uthi ollama pull [nombre-del-modelo]noma ngabe usebenzisa ollama run, uhlelo izolanda imodeli ngokuzenzakalelayo uma ungakabi nayo.
Kunezinhlobo ezahlukene zamamodeli kuye ngezidingo zakho:
- Ingxoxo: I-Llama 3 noma i-Mistral yizona zinkosi lapha ngenxa yokulinganisela kwazo phakathi kwekhwalithi kanye nesivinini.
- Ukuhlela: I-CodeLlama noma i-DeepSeek-Coder zilungele ukulungisa amaphutha ekhodi noma ukukhiqiza imisebenzi.
- I-Multimodal (Umbono): Amamodeli afana ne-LLaVA akuvumela ukuthi ulayishe izithombe bese ucela i-AI ukuthi izichaze.
- Ukuzindla (Ukucabanga): Amamodeli aklanyelwe ukuchaza izinqubo isinyathelo ngesinyathelo.
Ukuze uhlanganyele, mane usebenzise ollama run llama3 futhi uzofaka umyalo osebenzisanayo. Ngaphakathi kwalesi sikhathi, ungasebenzisa imiyalo efana nokuthi /? ngosizo noma /bye Ukuze uphume. Uma ufuna ukubona ukuthi yini oyifakile, ollama list Izokunikeza uhlu oluphelele, kanye ollama ps Ungahlola ukuthi imodeli iyasebenza yini kulayishwe ku-GPU noma ku-CPU.
Izilungiselelo ezithuthukisiwe kanye nokwenza ngezifiso
Uma ungumthuthukisi, i-Ollama iyimayini yegolide ngoba iveza i -REST API ku-port 11434. Lokhu kukuvumela ukuthi uxhume i-AI yendawo kunoma yiluphi uhlelo lokusebenza. Iyahambisana nefomethi ye-OpenAI, ngakho ungayihlanganisa ku-VS Code nge-GitHub Copilot (usebenzisa inketho ye-BYOK) noma usebenzise ama-ejenti afana ne-OpenCode.
Esinye isici esinamandla yi- Modelfile . Ifana ne-Dockerfile kodwa i-AI. Ikuvumela ukuthi udale inguqulo eyenziwe ngokwezifiso yemodeli ngokuchaza i -System Prompt ethile (isibonelo, ukuguqula i-Llama ibe uchwepheshe emthethweni waseSpain), ukulungisa izinga lokushisa ukuze lenze libe nobuciko noma libe nembe kakhudlwana, futhi ulondoloze lokho kulungiselelwa ngaphansi kwegama elenziwe ngokwezifiso.
Kulabo abafuna isikhombikubona esibonakalayo esifana kakhulu ne-ChatGPT, sincoma ukufaka i-Open WebUI . Ixhuma ku-Ollama njenge-backend futhi inikeza ulwazi oluphelele lwewebhu ngomlando wengxoxo kanye nokuphathwa kwedokhumenti, konke kusebenza kunethiwekhi yakho yendawo.
Amathiphu okuthuthukisa nokusebenza kahle
Uma uqaphela ukuthi i-AI isebenza kancane, isinyathelo sokuqala ukuhlola inani . Le nqubo yehlisa ukunemba kwezisindo zemodeli (kusukela kuma-bits angu-16 kuya kuma-bits angu-4 noma angu-8, isibonelo), okunciphisa kakhulu i-RAM edingekayo ngaphandle kokulahlekelwa ikhwalithi eningi kakhulu. Imodeli ye -Q4_K_M ivame ukuba yindawo enhle phakathi kokusebenza nokunemba.
Ukuze uvimbele i-Ollama ekusebenziseni izinsiza uma ingasetshenziswa, ungakhubaza ukuqala okuzenzakalelayo ku-Windows Task Manager. Kuyasiza futhi ukuqapha ukusetshenziswa kwe-VRAM ngesikhathi sangempela ukuze kutholakale ukuthi imodeli iyayikhipha yini i-RAM yesistimu, okunciphisa kakhulu ukusebenza.
Ukulawula ingqalasizinda yendawo kwenza ukusetshenziswa kobuhlakani bokwenziwa kube lula, kususwe izithiyo zomnotho zokubhalisela kanye nezingozi zokuphepha zefu. Ngokuhlanganisa amandla amamodeli afana ne-Llama 3 noma i-Mistral kanye nokuphathwa kalula kwe-Ollama, noma yimuphi umthandi noma inkampani ingakha uhlelo lwayo lwe-AI oluyimfihlo , ithuthukise ihadiwe etholakalayo futhi yenze ngokwezifiso ukuziphatha kwemodeli ngama-Modelfiles nama-API ahlanganisiwe.