Ikhono lokusebenzisa ubuhlakani bokwenziwa kumaseva ethu liye lasuka ekubeni yiphupho lomuntu othanda izinto eziningi laya ekubeni yisidingo esibalulekile. Namuhla, ubukhosi bedijithali kanye nobumfihlo kuyizinto eziqhuba izinkampani ukuba zisuke kuma-API ezentengiselwano futhi zakhe izimiso zazo zemvelo zamamodeli olimi, okuvimbela idatha ebucayi ukuthi igcine emafwini ezinkampani zangaphandle.
Ukuze kufezwe lokhu, amathuluzi afana ne-Ollama asebalulekile , esebenza njengebhuloho elilula lokuphatha ama-LLM ngaphandle kwezinkinga. Akukhona nje ukulanda imodeli nokuxoxa, kodwa mayelana nokuqonda ukuthi ungavumelanisa kanjani lobo buchwepheshe nezidingo ezithile, kungaba ngokwenza ngcono ihadiwe noma ukuqeqesha i-AI ngedatha yethu ukuze ikhulume ulimi lwebhizinisi lethu.
Izisekelo ze-AI Yendawo kanye ne-Ollama
I-Ollama empeleni iyiklayenti elisiza ekusetshenzisweni kwamamodeli olimi emshinini wakho. Isekelwe kulabhulali ye -llama.cpp , eyivumela ukuthi isuse ubunzima bobuchwepheshe futhi inikeze ulwazi olushelelayo. Enye yezinzuzo zayo ezinkulu ukuthi ivumela ukusebenza ngaphandle kokuxhumeka kwe-inthanethi , isuse noma yiluphi uhlobo lokuhlolwa kwangaphandle futhi iqinisekise ukuthi izicelo namadokhumenti akulokothi kuphume kunethiwekhi yenkampani.
Uma sikhuluma ngamamodeli amahhala, sisho lawo avumela ukufinyelela ezisindweni zemodeli futhi anamalayisense avulekile njenge-MIT noma i-Apache 2.0. Izibonelo ezivelele zifaka phakathi i-DeepSeek-r1 , efanelekile ekucabangeni kokuhlaziya; i-Llama 3.2 yokukhiqiza umbhalo ojwayelekile; i-Phi-4 ye-Microsoft yemisebenzi elula nephumelelayo; kanye ne -Mistral , elinganisela kahle ukulandela imiyalelo.
Isihluthulelo Sokusebenza Kahle: Ukulinganisa kanye Nehadiwe
Ukuze kufakwe imodeli enkulu kwikhompyutha ejwayelekile, kusetshenziswa i-quantization . Le nqubo ihilela ukunciphisa ukunemba kwezisindo zemodeli (kusuka kuma-bits angu-16 noma angu-32 kuya kuma-bits angu-4 noma angu-8), okunciphisa kakhulu usayizi wefayela futhi kwehlise kakhulu i-RAM edingekayo ngaphandle kokubeka engcupheni ikhwalithi yempendulo.
- I-RAM: Nakuba i-8GB yanele amamodeli amancane, ilungele kakhulu ukugeleza koketshezi ngamamodeli angu-7B noma angu-13B wakhe Imemori engu-16GB noma engu-32GB.
- I-GPU: Nakuba ungasebenzisa i-CPU, ukuba nekhadi lehluzo eline- i-VRAM eyanele Kungukushintsha umdlalo kwangempela, kusheshisa kakhulu ukuqagela.
- CPU: Amaprosesa esimanje asekelayo Imiyalelo ye-AVX512 ukuze kuthuthukiswe ukuphindaphinda kwe-matrix.
Ukuqeqeshwa Okuqondene Nomuntu: Kusukela ku-Mistral kuya ku-Custom Model
Uma amamodeli ajwayelekile ehluleka, isinyathelo esilandelayo ukulungisa kahle . Ukuhamba komsebenzi kobungcweti kuhilela ukusebenzisa ukwakheka kwe-Mistral-7B kuhlanganiswe ne-LoRA (Low-Rank Adaptation) kanye nethuluzi le-Axolotl. Le ndlela ivumela ukuqeqesha imodeli ngaphandle kokushintsha wonke amapharamitha ayo, konga isikhathi kanye namandla okusebenzisa ikhompyutha.
Inqubo iqala ngokuthi a isethi yedatha ehlelekile ngefomethi ye-JSONL noma ye-YAML, lapho izindima zichazwa khona njengokuthi system, user y assistantKuqeqesho, kuvamile kakhulu ukusebenzisa izindawo ezikude njenge-RunPod, enikeza ama-GPU e-A100 ngamanani angabizi, okukuvumela ukuthi usebenzise umyalo axolotl train config.yaml ukukhiqiza ama-adapter.
Uma isiqeqeshiwe, i-adaptha ye-LoRA kumele ihlanganiswe nemodeli eyisisekelo kusetshenziswa izikripthi ze-Python ukudala imodeli engaguquki. Okokugcina, ukuze i-Ollama iyifunde, kubalulekile ukuguqula umphumela ube yifomethi ye-GGUF , uyilinganise (isibonelo, ibe yi-q8_0) ukuze ihambisane nehadiwe yendawo.
Ukuthunyelwa Nokuphathwa Ezindaweni Zangaphakathi
Ukuze ufake imodeli ekukhiqizweni, inketho engcono kakhulu DockerNgefayela docker-compose.ymlSingaphakamisa isitsha se-Ollama ukufinyelela ngqo ku-GPU kanye namavolumu aqhubekayo ukugwema ukulahlekelwa amamodeli lapho eqala kabusha. Ukubhaliswa kokugcina kwenziwa ngokudala i- Ifayela lemodelilapho sichaza khona izinga lokushisa (ubuciko) kanye nomongo (num_ctx) ngaphambi kokusebenzisa ollama create.
Ukuqinisekisa ukuthi ulwazi aluyona nje isikrini esimnyama se-terminal, i-Open WebUI ihlanganisiwe . Lesi sixhumanisi esibonakalayo esinemifanekiso sikuvumela ukuthi uphathe abasebenzisi, udale amathempulethi asheshayo, futhi uxhumeke kuseva ye-Ollama usebenzisa ikheli le-IP kanye ne-port (ngokuvamile i-11434). Iyithuluzi elifanele kakhulu kubasebenzisi abangebona ochwepheshe ukuthi basebenzisane ne-AI yenkampani.
Ukusebenzisana Kwebhizinisi kanye Namacala Okusetshenziswa
Ezindaweni eziyinkimbinkimbi kakhulu zezinkampani, izingqimba zokuhlelwa kwemisebenzi njenge -SoaxNG esekelwe ku-OpenStack zingasetshenziswa . Lokhu kuvumela ukudalwa kwezinhlelo zemvelo ezihlanganisiwe ezisebenzisa ukuguquguquka kwamafu ngenkathi zigcina amakhono okuphetha endaweni . Lokhu kubalulekile emikhakheni efana nokunakekelwa kwempilo kanye nezezimali, lapho ukuthobela i-GDPR kanye ne-ISO 27001 kuyimpoqo.
Ezinye izinhlelo zokusebenza zangempela zifaka:
- I-Cybersecurity: Ukudalwa okuzenzakalelayo kwezincwadi zokudlala zempendulo yezehlakalo ezisekelwe ku-MITRE ATT&CK.
- I-DevOps: Ukuhlaziywa kwekhodi okuphephile kanye neziphakamiso zokulungisa ngokuzenzakalelayo.
- Kwezomthetho: Ukuqapha izinguquko zomthetho ngesikhathi sangempela kanye nokukhishwa kwedatha yenkontileka kusetshenziswa amamodeli okubona afana nalawa LLaVA.