Cishe uke wazibuza ukuthi kungenzeka yini ukuba nobuhlakani obunamandla bokwenziwa ngaphandle kokuthembela kumaseva enkampani enkulu e-United States ukuthi asebenze. Impendulo inguyebo ozwakalayo, futhi ngaphezu kwalokho, ukusebenzisa amamodeli olimi kudivayisi encane njenge-Raspberry Pi akuseyona nje indlela yokuhlola ama-geek; sekuyindlela yobuchwepheshe esebenzayo futhi ephumelelayo ngokumangazayo.
Lo mlingo wenzeka ngenxa ye-edge computing , okuhilela ngokuyinhloko ukucubungula ulwazi lapho lukhiqizwa khona. Ngokuhambisa i-AI kudivayisi, sifinyelela ubumfihlo bedatha obuphelele , njengoba kungekho lutho oluphuma ohlelweni, futhi siqeda ukubambezeleka okukhungathekisayo okwenzeka lapho idatha kufanele ihambe izinkulungwane zamakhilomitha ngaphambi kokuthola impendulo.
Ihadiwe efanelekile ye-AI yendawo
Uma ufuna ukuhlolisisa lokhu, akulona nje noma yiliphi ibhodi elizosebenza. I -Raspberry Pi 5 iyindlela ephelele yokusekela ngenxa yeprosesa yayo ye-Arm Cortex-A76, enamandla amakhulu okusingatha imisebenzi yokuqagela. Ukuqinisekisa ulwazi olubushelelezi, kubalulekile ukuba nemodeli ye -RAM engu-8GB , njengoba amamodeli alinganisiwe adla inani elikhulu le -RAM , futhi uhlelo lokusebenza ludinga isikhala esanele sokusebenza.
Indawo eyodwa lapho abantu abaningi benza khona amaphutha izinga lokushisa. Ukuqagela kwe-LLM kusunduza ama-cores ku-100%, okwenza iprosesa ishise ngokushesha kakhulu. Ukuze kuvinjelwe ukwehla kokusebenza kwesistimu ngenxa yokushisa , i -Active Cooler esemthethweni akuyona into ongayikhetha; iyimpoqo. Uma ungafuni i-Raspberry Pi yakho iphenduke i-toaster, udinga lokho kupholisa okusebenzayo.
Ngokuphathelene nesitoreji, yize ikhadi le-Class A2 microSD lizokwanela, uma ufuna amamodeli alayishe ngokuphazima kweso, ikhambi elifanele ukusebenzisa i- NVMe SSD nge-M.2 HAT. Umehluko mkhulu kakhulu: ukulayisha imodeli engu-2GB kungasuka kumasekhondi angu-12 kuya ku-3 noma 4 kuphela, okusheshisa ukuthunyelwa kwanoma yiluphi uhlelo lokusebenza onqenqemeni.
Ukuqonda ama-SLM kanye nokulinganisa
Khohlwa ngokuzama ukusebenzisa i-GPT-4 ku-Raspberry Pi; kungaba njengokuzama ukufaka indlovu emotweni encane. Yilapho kungena khona ama-Small Language Models (SLMs) . Lawa mamodeli, ngokuvamile anamapharamitha aphakathi kwezigidi ezimbalwa kanye nezigidi eziyi-7 noma eziyi-8, aklanyelwe ngqo amadivayisi anezinsiza ezilinganiselwe ngaphandle kokulahlekelwa ukuhambisana okuningi.
Isihluthulelo sokwenza lokhu kusebenze yi -GGUF quantization . Empeleni, kuhilela ukunciphisa ukunemba kwezisindo zemodeli (isibonelo, kusukela kuma-bits angu-16 kuya kuma-bits angu-4). Lokhu kwenza imodeli ithathe i-RAM encane kakhulu futhi igcina isivinini sokukhiqiza amathokheni samukelekile, okuvumela isivinini sokufunda esikhululekile kumuntu.
- I-Llama 3.2 (1B kanye no-3B): Kuhle kakhulu ezingxoxweni zezilimi eziningi kanye nemisebenzi yokufingqa.
- I-Gemma 3 kanye ne-3n: Zivelele ngokusebenza kwazo kahle, futhi kwezinye izinguqulo, zivelele ngamakhono okubona.
- I-Microsoft Phi-3.5: Unamandla kakhulu ekucabangeni, yize ngezinye izikhathi ekwazi ukukhuluma kakhulu.
- I-TinyLlama: Inkosi yesivinini, ilungele imiyalo elula yokwenza izinto ngokuzenzakalela ekhaya.
Amathuluzi okusebenzisa: i-Ollama ne-Llama.cpp
Ukuze konke lokhu kuqale ukusebenza, sinezindlela ezimbili eziyinhloko. Ngakolunye uhlangothi, i-Llama.cpp iyindlela yokukhetha kulabo abafuna ukulawula okuphelele. Ikuvumela ukuthi uhlanganise ikhodi yomthombo, wenze ngcono imiyalelo ye-ARM NEON kanye ne-dotprod, ecindezela yonke ithonsi lokugcina lamandla kusuka ku-silicon. Ilungele ukudalula i- API ehambisana ne-OpenAI ku-port 8080 nokuxhuma i-Pi kwamanye amadivayisi kunethiwekhi yendawo.
Ngakolunye uhlangothi, sine -Ollama , okungenzeka ukuthi iyindlela elula yokuphatha amamodeli namuhla. Ngemiyalo embalwa ku-terminal, ungalanda futhi usebenzise amamodeli afana ne-Llama noma i-Gemma ngaphandle kokuphazamiseka. I-Ollama iqala iseva ngemuva evumela ukuthi uxhumane ne-AI ngokusebenzisa umtapo wolwazi we-Python onembile kakhulu , okwenza kube lula ukudala izikripthi ezenziwe ngokwezifiso.
Ukuze uthuthukise uhlelo, ikakhulukazi ezinguqulweni ze-Lite ze-Raspberry Pi OS, kubalulekile ukwandisa isikhala sokushintshana sibe cishe yi-4GB. Lokhu kuvimbela inqubo ukuthi iqedwe yi-OOM Killer (Out Of Memory) lapho imodeli yengxoxo kanye nomongo kuqala ukugcwalisa i-RAM etholakalayo.
Amamodeli Ombono Nolimi (ama-VLM) asemaphethelweni
Izinto ziba mnandi kakhulu uma sihlanganisa ukubona nenkulumo. Amamodeli Olimi Lokubona (ama-VLM) avumela i-Raspberry Pi ukuthi ingagcini nje ngokufunda, kodwa nokuqonda izithombe. Amamodeli afana ne -Moondream ashesha ngokumangalisayo futhi ayakwazi ukuchaza izigcawu, ukubala izinto, noma ukwenza i-OCR ngqo kudivayisi.
Isibonelo esisebenzayo nesinamandla ukusetshenziswa kwe- Raspberry Pi AI Camera . Esikhundleni sokuthumela ividiyo eluhlaza efwini, ikhamera icubungula isithombe esisenzwa bese ikhiqiza imethadatha (njenge-object tag kanye namazinga okuzethemba). Le datha elula ithunyelwa ku-LLM, eyiguqula ibe yizifinyezo ezifundeka kalula . Lona umehluko phakathi kokuthumela ividiyo engu-1GB kanye nefayela lombhalo elingu-1KB.
Le ndlela ivula umnyango wezinhlelo zokusebenza ezifana nokuqapha ishelufu lokuthengisa , lapho uhlelo luxwayisa khona uma isitokwe singekho, noma ukubhekwa kwefektri ukuhlola ukuthi opharetha bagqoke yini imishini yokuphepha , konke lokhu ngenkathi begcina ubumfihlo futhi behambisana nemithetho ye-GDPR ngokungalayishi izithombe kumaseva angaphandle.
Amacala okusetshenziswa kwangempela kanye nokwenza ngokuzenzakalela
Uma uthanda ukwenza izinto ngokuzenzakalela ekhaya, ungaguqula i-Raspberry Pi yakho ibe yisikhungo sokulawula sendawo . Cabanga ngomsizi wezwi osebenzisa i-Whisper.cpp ukuloba izwi lakho kanye ne-LLM yendawo ukuhlaziya inhloso ibe yi-JSON, bese kuba nomthelela esenzweni ku-Home Assistant. Konke lokhu kwenzeka ngama-millisecond futhi ngaphandle kokuxhumeka kwe-inthanethi.
Ezindaweni zezimboni, lezi zinhlelo zingasetshenziswa ekulungiseni kusengaphambili noma ekwenzeni ngcono inqubo ngesikhathi sangempela. Zibaluleke kakhulu kwezolimo ezinembile, lapho idivayisi yeselula ingahlaziya impilo yezitshalo ensimini ngaphandle kokudinga ukumbozwa kwe-5G.
Kusukela kumathuluzi okufundisa ezindaweni ezikude kuya kubasizi babantu abakhubazekile, ikhono lokusebenzisa i-AI ekhiqizayo emaphethelweni lenza ukufinyelela kobuchwepheshe kube lula futhi livumela ukudalwa kwezixazululo ezikhethekile kakhulu ezingaxhomekile emalini yokubhalisa yamafu yanyanga zonke.
Ukuba nekhono lokucubungula ulimi kanye nombono endaweni ebhodini elibiza ama-€80 kuyisinyathelo esikhulu kwezobuchwepheshe. Ngokuhlanganisa ihadiwe elungiselelwe kahle, amamodeli alinganisiwe, namathuluzi afana ne-Ollama, noma yimuphi unjiniyela angakha uhlelo lwe-AI olusheshayo, olusebenzayo, noluyimfihlo oluwusizo ngempela emhlabeni ongokoqobo.
