- Ukwehlukaniswa kobuchwepheshe phakathi kwamamodeli anesisindo esivulekile kanye nomthombo ovulekile wangempela kubuhlakani bokwenziwa.
- Ukufana okunesu phakathi kokwamukelwa ngobuningi kwama-Kubernetes ezitsheni kanye nomkhuba wamanje wamamodeli avulekile.
- Ukusetshenziswa okusebenzayo kwezakhiwo zokuphetha ezilungiselelwe kusetshenziswa i-vLLM kanye ne-KubeAI ezindaweni zamafu.
- Umthelela we-geopolitical kanye nowezomnotho wokubusa ngentando yeningi kwesisindo samamodeli uma kuqhathaniswa nokulawulwa kwama-laboratory avaliwe.
Uma sibheka emuva ku-2015, noma ubani owayefuna ukusetha uhlelo olusatshalaliswe wabhekana nenkinga. Kwakukhona i-Apache Mesos, eyayisivele isunguliwe futhi ikhethwa kakhulu yiziqhwaga ezifana ne-Twitter ne-Airbnb, kanti ngakolunye uhlangothi, i-Docker Swarm, eyayilula kakhulu futhi ijwayelekile. Phakathi kwakho konke lokhu, kwavela umuntu omusha ogama lakhe linguKubernetes, owasungulwa yi-Google. Ngaleso sikhathi, ukuhlakanipha okwakukhona kwakuwukuthi iMesos yayiyeyengqalasizinda yangempela nokuthi iKubernetes yayiyinto yokudlala nje. Ngisho ne-Amazon yanquma ukuqalisa i-ECS yayo esikhundleni sokugxumagxuma. Kodwa sonke siyazi ukuthi leyo ndaba yaphela kanjani.
I-Kubernetes ayizange iwine ngoba yayiyibuchwepheshe obuthuthuke kakhulu ngaleso sikhathi, kodwa ngoba yakwazi ukuba yisikhungo sokudonsa amandla embonini . Yaguquka yaba isisekelo esingathathi hlangothi lapho abahlinzeki bamafu, onjiniyela, nabathengisi babengakha khona ngaphandle kokwesaba. Lapho isifinyelele kuleyo nqwaba ebalulekile, ukusungula izinto ezintsha kwaqhuma: isitoreji, ukuphepha, kanye nokubonakala kwaqala ukuxazululwa ngenxa yomphakathi. Namuhla, sibona uhlelo lwe-artificial intelligence luphinda umbhalo ofanayo, futhi labo abaqonda leli phethini bazokwazi ukwenza izinqumo zobuchwepheshe ezinolwazi oluningi.
Ama-Peso Avulekile noma Umthombo Ovulekile? Azifani
Ukuze sigweme ukudideka, ake sicacise eminye imiqondo. Abantu abaningi babiza amamodeli ngokuthi "umthombo ovulekile" uma empeleni e -open-weighted . Lokhu kusho ukuthi ungalanda amapharamitha aqeqeshwe kusengaphambili, uwalungise, bese uwasebenzisa noma kuphi lapho ufuna khona, kodwa awunakho ukufinyelela kudatha yokuqeqesha noma yonke inqubo yokudala. I-Open Source Initiative (OSI) iqine kakhulu: kubo, i-AI evulekile kumele ifake ikhodi yokuqeqesha kanye nesethi yedatha esetshenzisiwe.
Kummeli, lo mehluko ubalulekile, kodwa unjiniyela ojwayelekile akanandaba uma nje ithuluzi lisebenza futhi likwazi ukwenziwa ngendlela oyifisayo. Kufana nokuqhathanisa ama-Kubernetes (umthombo ovulekile ngokuphelele) nokusatshalaliswa kwe-binary Linux; uthola into ehlanganisiwe futhi ungayishintsha, noma ngabe ipayipi lokwakha lokuqala liphethwe ngumdali. Ekugcineni, umphakathi ubeka phambili ukusebenziseka ngaphezu kobumsulwa belayisensi, ucabangela izici ezifana nomthwalo wemfanelo ekuhlakanipheni kokwenziwa kanye nezinselele zayo zokuziphatha.
I-ecosystem isivele ilapha futhi ihamba ngesivinini esikhulu.
Isivinini esikhula ngaso lesi simo siyamangalisa. I-HuggingFace isivele isingatha izigidi zamamodeli, futhi eduze kwemindeni efana ne-Llama, i-Mistral, i-Qwen, ne-Gemma, konke okucatshangwayo kuyathuthukiswa: kusukela ezinguqulweni ezilinganiselwe zokusebenza kumadivayisi eselula noma i-Apple Silicon, kuya kuma-adaptha e-LoRa akhethekile kwezomthetho, ezokwelapha, noma izinhlelo. Ngaphezu kwalokho, kuye kwavela izikhathi zokusebenza ezifana ne-vLLM ne-SGLang ezilawula ukusebenza okuphezulu ngokuhlanganisa okuqhubekayo, kuyilapho i-Ollama ikuvumela ukuthi uqalise imodeli endaweni ngomyalo owodwa.
Kwakukhona isikhathi lapho impikiswano ngokumelene namamodeli omthombo ovulekile yayiwukuthi ayengakwazi ukuncintisana ne-GPT-4 noma i-Claude. Kodwa-ke, leso sikhala sesicishe sivalwe ngokuphelele. Amamodeli afana ne-GLM-5.2 noma i-Kimi K3 abonisa ukusebenza okuphambili , ikakhulukazi emisebenzini yekhodi eyinkimbinkimbi, ngezinye izikhathi enza kahle kakhulu kunezinguqulo zomthombo ovaliwe kumabhentshimakhi athile. Lapho amamodeli omthombo ovulekile "emahle ngokwanele," umphumela wenethiwekhi owashukumisa ama-Kubernetes uqala ukusebenza ngamandla angenakuvinjelwa.
Ukufana okuqondile: Kusukela ezitsheni kuya ku-AI
Uma sihlaziya isakhiwo, ukufaniswa cishe kunembile. Amamodeli ayisisekelo (Llama, Qwen) asebenza njenge-Docker ye-AI: ahlinzeka ngendawo yokuqala ejwayelekile noma yimuphi unjiniyela angayilanda futhi ayihlele ngokwezifiso, njengoba nje senze ngezithombe ze-Ubuntu noma ze-Alpine. Okwamanje, amathuluzi afana ne-Ollama noma i-llama.cpp agcwalisa umsebenzi we-Docker Compose, okwenza ukuhlanganisa imodeli endaweni yokuthuthukiswa kwendawo kube lula njengokwengeza isitsha se-PostgreSQL.
Isinyathelo esilandelayo ungqimba lwezindinganiso, olufana ne-Kubernetes. Nakuba lusachazwa, sesivele sibona izingcezu: amafomethi e-GGUF noma e-GPTQ asebenza njengezithombe ze-OCI, i-API ehambisana ne-OpenAI iyisixhumi esijwayelekile, kanti i-Hugging Face iyi-Docker Hub yamamodeli. Noma ngubani okwaziyo ukuba yingcweti yalolu ngqimba lwesevisi nokusabalalisa uzobamba iningi lezinto ezintsha embonini.
Ukuqaliswa okusebenzayo eKubernetes
Kulabo abasebenza ngeJava kanye neSpring Boot, lesi yisikhathi esibalulekile. Ngenxa yezinhlaka ezifana neSpring AI kanye neLangChain4j, manje kungenzeka ukuthuthukisa ngokumelene nemodeli yendawo bese uthuthela eqenjini lokukhiqiza ngokushintsha nje isici kufayela lokucushwa. Asisathembeli kukhiye be-API noma idatha yangaphandle ephuma kunethiwekhi yethu, okubalulekile emikhakheni efana nebhange kanye nokunakekelwa kwempilo lapho ubumfihlo bedatha bubaluleke kakhulu.
Ngokombono wobuchwepheshe, kunezindlela ezimbili eziyinhloko zokusebenzisa ama-Kubernetes (ikakhulukazi ku-GKE). Ngakolunye uhlangothi, singasebenzisa i-vLLM ngqo njengenjini yokucabanga ukuze sithole ukulawula okuphezulu kokusebenza. Ngakolunye uhlangothi, singakhetha i-KubeAI, ipulatifomu yendabuko yama-Kubernetes yokuphathwa kwamamodeli. I-KubeAI ikuvumela ukuthi uphathe ikhathalogi yamamodeli futhi inikeza izici ezifana ne -scale-to-zero , okunciphisa izindleko zokusebenza ngokungagcini ama-GPU evuliwe uma kungekho zicelo, yize iveza ukubambezeleka kokuqala okubandayo.
Impikiswano yezomnotho kanye neyezwe
Akusikho konke ukuqiniseka kwezobuchwepheshe; kukhona impi ebandayo eqhubekayo. Amamodeli aseShayina athola indawo emangalisayo ekulandeni, okwenza eminye imikhakha e-US icabangele imikhawulo. Kodwa-ke, cishe akunakwenzeka ukuvimbela imodeli ngokusekelwe emvelaphi yayo, njengoba izisindo zimane ziyizinombolo futhi azinalo ilebula lobuzwe. Noma yimuphi umzamo ongenangqondo wokuvinjelwa kungaba lula ukuwugwema.
Ngaphezu kwalokho, kukhona ukungezwani komnotho. Abanye ochwepheshe bathi amamodeli okulinganisa isisindo avulekile "awathandi ukwehla kwesidingo" ngoba, ngokunciphisa inani elingabanjwa amalabhorethri asemngceleni, angavimbela ukutshalwa kwezimali okukhulu kwengqalasizinda (i-CAPEX). Uma ukutshala imali eyizigidigidi ezingama-700.000 zamaRandi kungaqinisekisi ukuthi kuzoba noyedwa wenzuzo, imali ingasuswa. Kodwa-ke, umlando usitshela ukuthi ukulinganisa okuvulekile kuvame ukusheshisa ukwamukelwa ngobuningi, kunciphisa izindleko zokungena ezinkulungwaneni zezinkampani ezintsha.
Uma ungumthuthukisi futhi ungafuni ukusala ngemuva, indlela ekahle ukuqala ukuzama ngamamodeli alinganiselwe endaweni. Awudingi i-GPU enkulu, njengoba amafomethi afana ne-Q4 evumela imodeli ye-7B ukuthi isebenze ngokwamukelekayo kuma-CPU anamuhla. Kubalulekile ukusebenzisa izixhumi ezisebenzisana ne-OpenAI , njengoba kuyindinganiso ye-de facto, kungakhathaliseki ukuthi usebenzisa i-vLLM, i-SGLang, noma i-LocalAI. Okokugcina, ukuqonda umehluko phakathi kwamafomethi e-quantization (njenge-Q4_K_M noma i-Q8_0) kuzokuvumela ukuthi uthuthukise ukusetshenziswa kwe-RAM kanye nokuphendula kwezinhlelo zakho zokusebenza.
Umlando wekhompyutha usifundise ukuthi amapulatifomu avulekile avumela ukwenziwa ngokwezifiso okukhulu ekugcineni adlula noma yimuphi umthengisi ovaliwe, kungakhathaliseki ukuthi izinsiza zomthengisi zinjani. Njengamanje sibhekene nenkathi yobuhlakani bokwenziwa be-Kubernetes, lapho ikhono lokusebenzisa amamodeli ngokwezifiso engqalasizinda elawulwayo libuyisela ubukhosi bezobuchwepheshe kubathuthukisi namabhizinisi.