Ukwenza kahle i-LLM yakho: Umhlahlandlela wesinyathelo ngesinyathelo
Ukwenza kahle i-LLM yakho: Umhlahlandlela wesinyathelo ngesinyathelo Ukungena kwakho kokugcina emhlabeni wobuhlakani bomuntu owenziwe ngezifiso. Uma ukhathazekile ngobumfihlo bamathuluzi we-AI asuselwa kumafu afana ne-chatgpt noma i-bard, awuwedwa. Intshisekelo ekusebenzeni kwangasese, amamodeli amakhulu olimi lwasendaweni (LLMS) ikhuphuka ngokushesha – futhi ngenxa yezizathu ezinhle: Ubumfihlo bedatha engcono, ukulawulwa okugcwele kokuphuma, futhi akukho ukuxhumana kwe-inthanethi okudingekayo. Cabanga nje ubuza imibuzo enamandla ye-AI ngaphandle kokuthumela idatha efwini. Lo mhlahlandlela uzokuhamba ngokusetha eyakho i-LLM, noma ngabe awuyena umthuthukisi noma ubuchwepheshe be-tech. Silungele ukuvula amandla omsizi wakho wangasese we-AI? Ake siqale.
Funda futhi: Qalisa eyakho i-AI Chatbot endaweni yangakini
Kungani ugijime i-LLM endaweni?
Kunezinzuzo ezibalulekile zokusingatha imodeli yakho yolimi olukhulu. Kokukodwa, kukubeka ekuphathweni kwedatha yakho. Amathuluzi wezentengiso we-AI asebenza kumaseva akude efu, okusho ukuthi okokufaka kwakho-akunandaba ukuthi uzwela kangakanani-amaseva avela eceleni. Ukugijima imodeli emshinini wakho kususa ubungozi.
Esinye isizathu izindleko. Imali yokubhalisa yokufinyelela ezingeni le-PRO-Izinga ku-AI APIs ingangeza isikhathi esidlule. Ukusingatha imodeli yasendaweni, ngenkathi kudinga ukusetha kokuqala ne-Hardware, kungasusa amacala aqhubekayo.
Ijubane futhi liyisici. I-LLM yendawo ayithembeli ekuxhumaneni kwe-inthanethi, okwenza ilungele imisebenzi ezindaweni ezikude noma ngesikhathi sokuphuma. Onjiniyela, ababhali, abacwaningi, kanye nama-hobbyists ngokufanayo baphendukela kule ndlela ukuze kube lula futhi basebenze.
Futhi funda: amakhono ayi-7 abalulekile okufanele akwazi ukwenza kahle i-2025
Ukukhetha imodeli efanelekile yezidingo zakho
Akuwona wonke ama-LLMS adalwe ngokulinganayo. Ngaphambi kokungena ekusethweni, kubalulekile ukuhlola ukuthi hlobo luni lwemisebenzi olindele ukuthi imodeli yakho iyenze. Amanye amamodeli ahloselwe usizo lokuxoxa, amanye okuqedwa kwekhodi noma ukufinyelelwa kwedokhumenti.
Ngokusetshenziswa okujwayelekile, imodeli evulekile ethandwa kakhulu namuhla yi-llama ka-meta (imodeli enkulu yolimi meta ai). I-Varionts yayo ehlukahlukene Uzothola nezinto zokutholwa njenge-alpaca, viluna, kanye nephutha elilungiselelwe imisebenzi ethile.
Amafayela wemodeli ajwayele ukwabiwa ku-inthanethi ngamafomethi ahlukahlukene anjenge-GGUF (ifayela elikhiqizwayo elikhiqizwayo), elenzelwe ukusebenza kahle kwememori. Lawa mafayela angasukela ngaphansi kwe-2GB kuya ngaphezulu kwe-30GB kuya ngobunzima. Khetha ngobuhlakani ngokususelwa kumandla akho wehardware nokusebenza okuhlosiwe.
Funda futhi: Faka i-LLM kuma-macos kalula
Ukufaka isoftware ekhiye: llama.cpp ne-ollama
Ukugijima i-LLM kudinga isoftware ekhethekile. Phakathi kwamathuluzi asebenziseka kalula futhi asebenza kahle atholakalayo namuhla yi-llama.cpp, ukusetshenziswa kwe-C ++ okwenzelwe amamodeli we-Llama kwi-CPUS ebangeni lebanga lomthengi.
Izinyathelo zokufaka ngokuvamile ziqondile:
- Landa futhi ufake ukwakheka kwakamuva kwe-LLama.cPP kusuka kumthombo we-github othembekile.
- Thola ifayela lemodeli elihambisanayo (kunconywe ngefomethi ye-GGUF) kusuka ku-Hub eqinisekisiwe yemodeli eqinisekisiwe njengokuqabulayo kobuso noma i-bloke.ai.
- Faka ifayela le-gguf kufolda eqokiwe ye-LLAMA.CPP.
Ungangena kwimodeli usebenzisa umugqa womyalo noma imibhalo esebenza ngokusebenzisana. Lokhu kusetha kukuvumela ukuthi uxoxe ngokuqondile nemodeli yakho oyikhethile ngaphandle kokuhileleka kweseva yangaphandle.
Kwabasebenzisi beMac besebenzisa i-Apple Silicon (M1, M2 Chips), iLlama.CPP isebenza kahle ngenxa yokwenza i-hardware yendabuko. Kulabo abakhululeke kangako basebenzisa indawo yokuhlala yesikhumbuzo, i-ollama yindlela enobungane. Inikeza isikhombimsebenzisi sokuqhafaza futhi isekela amafomethi afanayo amamodeli wokusetha okusheshayo.
Futhi funda: I-NVIDIA iqala amamodeli amasha we-LLM we-AI
Ukwenza kahle isivinini nokusebenza
Ngenkathi ama-desktops aphezulu aphezulu ane-GPUs eqinile anikela ngokusebenza okuhle kakhulu, ama-LLM anamuhla aya ngokuya elungiselelwe ukusetshenziswa kwe-CPU. I-Llama.cPP isebenzisa amamodeli ahlukanisiwe, okusho ukuthi ukucaciswa kwezibalo kuncishisiwe ezindaweni ezingabucayi ukwenza ngcono ukucubungula ijubane ngaphandle kokulahlekelwa yikhwalithi.
Ngemiphumela emihle kakhulu, hlangana nalezi zinto ezilandelayo:
- Ubuncane be-8 GB RAM (16 GB ilungele)
- I-Apple Silicon M1 noma entsha (yabasebenzisi beMac)
- I-Quad-Core Intel noma i-AMD CPU (yabasebenzisi beWindows / Linux)
- I-SSD enikezelwe ngokulayisha ngokushesha
Kusetshenziswa izinhlobo ezincane ezihlukaniswe ngamamodeli (ama-4-bit noma ama-5-bit) kungathuthukisa kakhulu isikhathi sokubulawa ngenkathi kugcinwa ukusebenza kwemisebenzi esejwayelekile njengokubhala okuyisisekelo noma ukufingqa kwedatha.
Ukuthuthukisa ukusebenza ngezandiso
Ukugijima i-LLM ngokwakhe kunamandla, kepha ungathatha amandla okusebenzisa izandiso. Abanye abathuthukisi bakha ama-Wrappers noma ama-plugins ukuxhuma i-LLMS ngamathuluzi anjengeziphequluli zewebhu, abafundi be-PDF, noma amaklayenti e-imeyili.
Izithuthukisi ezijwayelekile zifaka:
- Imemori Yemezwi: Gcina umlando wokuxhumana futhi uvumele imodeli ukukhumbula imiyalo edlule
- Inkulumo-to-Umbhalo: Guqula Imiyalo Yezwi ibe yimodeli
- I-APIS: Izicelo zangaphandle zangela amakhalenda noma imininingwane yolwazi
Lawa ma-plugins avame ukudinga amakhono okuhlela okukhanyayo ukufaka futhi wenze ngezifiso, kepha abaningi beza nama-tutorials kanye nemibhalo ukwenza lula ukusetshenziswa.
Ukuhlala ngasese futhi kuphephile
Enye yezizathu eziphambili zokusetha i-LLM yendawo ukuqinisekisa ubumfihlo. Lokho akusho ukuthi ungaphumula ukuma kwakho kwezokuphepha. Gcina i-laptop yakho noma ideskithophu yakho ivikelwe isoftware ye-antivirus futhi uvuselele uhlelo lwakho lokusebenza njalo ukukhawula ubungozi.
Landa kuphela amafayela wemodeli nemibhalo yokusetha evela emithonjeni ethembekile. Ukuqinisekiswa kokuhlola ukuqinisekisa ukuqinisekisa ukuthi amafayela awashintshiwe. Uma usebenzisa ama-wrappers noma ama-plugins ngokwezifiso, buyekeza ikhodi yomthombo ngokwakho noma uthintane nezinkundla zomphakathi ukuqinisekisa ukuphepha.
Ukusetshenziswa okungaxhunyiwe ku-inthanethi kuyisiqinisekiso sakho esihle kakhulu sobumfihlo. Lapho imodeli ilandwe bese isetha, kufanele ukwazi ukunqamula kwi-Intanethi futhi uqhubeke nokusebenzisa i-LLM yakho ngaphandle kwendaba.
Izeluleko Ezijwayelekile Zokuxazulula Izinkinga
Noma ngokulungiselela okuhle, ungahle ushaye ama-snags ngezikhathi ezithile ngesikhathi sokufakwa noma ukwenziwa kwemodeli. Ezinye izindaba ezivamile zifaka:
- “Amaphutha angekho emthethweni”: Lokhu kuvame ukwenzeka uma i-CPU yakho ingasekeli isethi yokufundisa esetshenziswe ngesikhathi sokuhlanganiswa. Zama ukulanda enye eyakha.
- Imithwalo yemodeli kepha ngeke iphendule: Lokhu kuvame ukuvela ekusebenziseni ifomethi yemodeli engalungile. Qinisekisa ukuthi usebenzisa i-GGUF noma okuhlukile okusekelwa.
- Izikhathi zokuphendula ezihamba kancane: Shintshela kwimodeli ephansi ephansi, noma hlola ukuthi idivaysi yakho ayisebenzi izinhlelo ezingemuva ezinamandla.
Bheka imiphakathi yabasebenzisi kwi-Reddit noma izingxoxo ze-GitHub ngezixazululo ezisheshayo. Iningi lalawa mapulatifomu manje afaka abasebenzisi abasebenzayo ababelana ngezimpendulo zesikhathi sangempela kanye namathiphu wokusetha.
Egijima amakhulu e-LLM
Ukuqalisa imodeli enkulu yolimi (LLM) kwikhompyutha yakho usebenzisa i-OLLAMA, landela umhlahlandlela wesinyathelo ngesinyathelo esingezansi. U-OLLAMA uhlaka olukuvumela ukuthi usebenze ama-LLMS ahlukahlukene endaweni yakini, njengamamodeli wesitayela se-GPT, emshinini wakho.
Izimfuneko:
- I-Mac noma iLinux (Ukuxhaswa KweWindows Kuza maduze)
- Izidingo ze-Hardware:
- Ikhompyutha okungenani I-8GB ye-RAM.
- Okungenani I-10GB yesikhala sediski samahhala amamodeli.
- Faka i-Docker (U-OLLAMA ugijima endaweni equkethe).
- Faka i-Docker kusuka lapha.
Isinyathelo 1: Faka u-Ollama
Ukufaka u-Ollama, landela le miyalo:
- Download ollama:
- Faka uhlelo lokusebenza:
- Phezu kwa- Inzivula
.dmgIfayela bese uhudula uhlelo lokusebenza lwe-Ollama kufolda yakho yezicelo. - Phezu kwa- Ilingu lelosebenzisa i-terminal ukufaka:
- Landela noma yiziphi izinyathelo zokusetha ezengeziwe kusuka kwisifaki.
- Phezu kwa- Inzivula
curl -sSL | bashIsinyathelo 2: Qalisa uhlelo lwe-Ollama
- Vula Ollama kusuka kwakho Ifolda Yezicelo ku-mac noma esigungwini se-linux.
- Bheka ukuthi i-ollama isebenza kahle:
- Vula i-terminal nohlobo: Lo myalo kufanele ubuyise uhlobo olufakiwe lwe-OLLAMA uma ukufakwa kuphumelele.
ollama --versionIsinyathelo 3: Qalisa imodeli nge-Ollama
U-OLLa uxhasa ukusebenzisa ama-LLMS ambalwa, njengamamodeli we-GPT. Ukugijima imodeli, sebenzisa izinyathelo ezilandelayo:
- Vula i-terminal:
- Vula isikhombimsebenzisi se-ukuphela noma umyalo kwikhompyutha yakho.
- Uhlu amamodeli atholakalayo:
- Ungabona ukuthi yiziphi amamodeli atholakala ngokugijima:
ollama models list
- Ungabona ukuthi yiziphi amamodeli atholakala ngokugijima:
ollama models list- Lokhu kuzokukhombisa uhlu lwama-llms atholakalayo ongawasebenzisa emshinini wakho.
- Gijima imodeli ethile:
- Ukuze usebenzise imodeli, ungasebenzisa:
ollama run - Beka endaweni
gpt-3nomachatgpt). - Gijima i-LLM kwimodi yokusebenzisana:
- Ukuqala iseshini exhumana lapho ungaxoxa khona nemodeli, uhlobo:
ollama run --interactive - Lokhu kuzovula ingxoxo esekwe kwi-terminal lapho ungathayipha khona imiyalezo, futhi imodeli izophendula ngokuhlanganyela.
Isinyathelo 4: Yenza ngokwezifiso ukuziphatha kwemodeli
Ungadlulisela amapharamitha athile ukuze wenze ngezifiso ukuziphatha kwemodeli. Isibonelo, ungaguqula izinga lokushisa (elilawula ubuhlakani), noma linikeze imiyalo ethile yezimpendulo ezilawulwa kakhulu.
- Setha amapharamitha:
- Isibonelo, ukulungisa amazinga okushisa, ungagijima:
ollama run --temperature 0.7 - Hlinzeka ngokusheshisa ngokwezifiso:
- Ungase futhi unikeze ngokuthuthukela ngokwezifiso kumodeli ekuqaleni. Ngokwesibonelo:
ollama run --prompt "Tell me about the future of AI." Isinyathelo 5: Hlangana namamodeli nge-API (Ngokuzithandela)
- Run ollama api:
- Uma ungathanda ukuhlanganisa imodeli ngekhodi yakho, ungasebenzisa i-Ollama's API. Ukuqala iseva ye-API:
ollama api start- Yenza izingcingo ze-API:
- Manje usungaxhumana nemodeli ngezicelo ze-HTTP, usebenzisa
curlNoma iyiphi ilabhulali yeklayenti le-HTTP kwikhodi yakho. Ngokwesibonelo:
- Manje usungaxhumana nemodeli ngezicelo ze-HTTP, usebenzisa
curl -X POST -H "Content-Type: application/json" -d '{"model": "", "prompt": "Hello, world!"}' Isinyathelo 6: Ukuqapha ukusetshenziswa kwezinsiza (kuyakhetheka)
Njengoba i-LLMS ingaba sezinsizakusebenza, ungabheka ukusetshenziswa kwezinsizakusebenza kohlelo lwakho ukuze kuqinisekiswe ukusebenza okubushelelezi.
- Qapha ukusetshenziswa kwe-CPU / Ram:
- Ku-mac, sebenzisa Ukuqapha Umsebenzi.
- Ku-Linux, sebenzisa:
top- Lungiselela ukusebenza:
- Uma imodeli ihamba kancane kakhulu noma izinsizakusebenza zohlelo lwakho zigcwele kakhulu, zama ukunciphisa inani lezinqubo ezisebenzayo noma ukuguqula usayizi wemodeli.
Isinyathelo 7: Ukuxazulula inkinga
- ISIHLOKO: Imodeli ayisebenzi:
- Uma imodeli ingalayishi, qiniseka ukuthi uhlelo lwakho luhlangabezana nezidingo ezimbalwa ze-Hardware nezidingo zesoftware. Bheka izingodo zamanye amaphutha usebenzisa:
ollama logs- ISIHLOKO: Ukusebenza kwemodeli kuphansi:
- Zama ukusebenzisa amamodeli amancane noma ukuvala ezinye izinhlelo zokusebenza zokukhulula izinsiza zohlelo.
Izinsizakusebenza ezengeziwe:
Isiphetho: I-AI yakho, imithetho yakho
Ukusetha imodeli yakho enkulu yolimi akusekho umsebenzi okhawulelwe ochwepheshe. Ngamathuluzi athuthukisiwe, amamodeli alungiselelwe, kanye nemihlahlandlela enemininingwane, noma ngubani angasebenzisa ngokunenzuzo abasizi bendawo be-AI. Noma ngabe ufuna ukuvikela idatha yakho, wonge imali, noma umane nje uzame ngomunye wezobuchwepheshe oguqukayo namuhla, ukusebenzisa i-LLM yakho yendawo kungukutshala imali okuhle. Landela lezi zinyathelo ukwethula isixazululo se-AI somuntu siqu esihlangabezana nezindlela zakho zobumfihlo nezidingo zokusebenza. Qala kahle i-LLM yakho namuhla futhi ulawule izingxoxo zakho zedijithali.
Ukunqubekela phambili
UParker, uProfesa uPhilip M., Ph.D. Umbono wezwe we-2025-2030 wezobunhloli bokufakelwa ekunakekelweni kwezempilo. Insual, 3 Mari. 2024.
Khang, Alex, Umhleli. Izinto ezintsha eziqhutshwa ngempilo yedijithali: izitayela ezivelayo, izinselelo kanye nezicelo. IGI Global, 9 Feb. 2024.
Singla, Babita, et al., Abahleli. Ukuguqula umkhakha wezempilo nge-AI. IGI Global, 26 Julayi 2024.
I-Topol, Eric J. Umuthi ojulile: Ukuthi ubuhlakani bokufakelwa bungenza kanjani umuntu wezempilo futhi. Izincwadi eziyisisekelo, ngo-2019.
UNelson, uJohn W., Umhleli, et al. Kusetshenziswa ama-analytics abikezelayo ukuthuthukisa imiphumela yokunakekelwa kwempilo. 1st ed., Heress, 2021.
I-Subbharaam, Vinithasree. Ukuhlaziya okubikezelayo ekunakekelweni kwezempilo, ivolumu 1: Ukuguqula ikusasa lezokwelapha. 1st Ed., Isikhungo Sokushicilela IPhysics, 2021.
Kumar, abhishek, et al., Abahleli. Kuvela ama-analytics okubikezela ekunakekelweni kwezempilo: amasu amasha we-AI wokungenelela kwangempela. Isikhungo sobunjiniyela nobuchwepheshe, 2022.
UTetteh, uHassan A. Ukunakekelwa kwezempilo ngobuhlakani nge-AI: Ukuhlanganisa umuthi wamasosha ukuguqula ukunakekelwa kwempilo kwawo wonke umuntu, yonke indawo. I-Forbesbooks, 12 Novemba 2024.
Umtholi, uTom. I-AI Empilweni: umholi womholi wokuwina e-New Age entsha yezinhlelo zezempilo ezihlakaniphile. 1st Ed., Hims, 13 Feb. 2020.
UHolley, uKerrie, noMathuru onesithunzi. I-LLMS kanye ne-ai ekhiqizayo yezempilo: umngcele olandelayo. 1st Ed., Emidiya ye-O'Iilly, 24 Septhemba. 2024.
UHolley, Kerrie, noSiupo Becker MD I-AI-First Healthcare: Izicelo ze-AI ekuphathweni kwebhizinisi nokuphathwa kwezempilo. 1st Ed., Emidiya ye-O'Iilly, 25 Meyi 2021.
