Maitiro ekuisa Gemma 3 LLM pa Windows 11 nhanho nhanho

Mienzaniso yemitauro yakavhurika yakashanduka nekusvetuka-svetuka, uye Nhasi zvinokwanisika kunakidzwa nazvo zvakananga kubva pakombuta yako pasina kutsamira pamasevhisi emakore.. Imwe yeanonyanya kuvimbisa pari zvino iGemma 3, Google's LLM nyowani yakavakirwa paGemini tekinoroji, iyo inomira pachena nekugona kwayo kugadzirisa zvinyorwa nemifananidzo, pamwe neiyo yakakura hwindo hwindo rekusvika 128k tokens mumashanduro ayo epamberi. Kuti uwane rumwe ruzivo nezve kuburitswa uku, unogona kushanyira chinyorwa chedu pa kuratidzwa kweGemma 3.

Kana uri kushandisa Windows 11 uye kutarisa kuisa Gemma 3 yekuedza kana kunyange kugadzirwa kwenzvimbo, mauya panzvimbo chaiyo. Ngationgororei zvakadzama nzira dzese dzinogoneka dzekuisimudza uye kushanda pakombuta yako, kusanganisira yakanyanya kukurudzirwa sarudzo seOllama, LM Studio, uye zvakare imwe-yakavakirwa gore neimwe nzira neGoogle AI Studio. Mukuwedzera, isu tichagadzirisa zvinodiwa zvehunyanzvi, zvakanakira imwe neimwe nzira uye Maitiro ekuita zvakanyanya kugona kweiyi simba rekugadzira njere.

Chii chinonzi Gemma 3 uye nei kuchiisa?

Gemma 3 chizvarwa chechitatu chemhando dzeLLM dzakaburitswa neGoogle pasi perezinesi rakavhurika-sosi.. Kusiyana nemhinduro dzakapfuura senge Llama kana Mistral, inopa rutsigiro rwakananga rwekuisa mufananidzo, mamiriro akafaranuka, uye rutsigiro rwemitauro inopfuura zana nemakumi mana. Zvichienderana nezvaunoda, unogona kusarudza kubva kune akati wandei mavhezheni emuenzaniso kubva pa140B kusvika 1B paramita:

• Gemma 3:1B: Yakareruka modhi yakanakira mabasa ekutanga uye zviwanikwa-zvishoma nharaunda.
• Gemma 3:4B: Kuenzanisa mashandiro uye kugona kwemashandisirwo epakati.
• Gemma 3:12B: Inokurudzirwa kuongororwa kwakaoma, kuronga uye mitauro yakawanda.
• Gemma 3:27B: Sarudzo ine simba kwazvo, yakagadzirirwa kushandiswa kwakanyanya, multimodal ine hukuru hwemamiriro ezvinhu.

Iko mukana wekumhanyisa yakadaro yemhando yepamusoro kubva kuPC yako Inoshandura mitemo yemutambo maererano nekuvanzika, kumhanya kwekupindura, uye kutsamira kune vechitatu mapato. Iwe hauchadi kubhadhara kunyoreswa kwemwedzi kana kusiya data rako. Chinongodiwa kugadzirira kushoma uye chido chekudzidza. Kana iwe uchida kuziva zvakawanda nezve mabhenefiti emhando idzi, tarisa chinyorwa chedu pa vhura uremu AI modhi.

Sarudzo yekutanga: Kuiswa naOllama

Ollama pamwe ndiyo nzira iri nyore yekumhanyisa LLM seGemma 3 kubva Windows 11. Yayo terminal-yakavakirwa interface inokutendera iwe kuti uise uye umhanye mamodheru ane yakapusa yekuraira mutsara. Uyezve, inoenderana neMacOS, Linux, uye Windows, zvichiita kuti zvive nyore kushandisa munzvimbo dzakasiyana siyana.

Matanho ekuisa Ollama uye kumhanya Gemma 3:

1. Svika webhusaiti yepamutemo: ollama.com.
2. Dhawunirodha installer yeWindows uye imhanye semamwe mapurogiramu.
3. Vhura Command Prompt (CMD) kana PowerShell uye simbisa kuisirwa ne:

ollama --version

Kana zvese zvikafamba zvakanaka, unogona ikozvino kurodha chero anowanikwa Gemma 3 matemplate. Ingomhanya imwe yemirairo iyi zvichienderana netemplate yaunoda:

ollama run gemma3:1b
ollama run gemma3:4b
ollama run gemma3:12b
ollama run gemma3:27b

Kana wangodhawunirodha, unogona kutanga modhi nyore nyore.. Kuti uite izvi, mhanya:

ollama init gemma3

Kubva panguva iyoyo zvichienda mberi, unogona kutanga kutaurirana neLLM ne:

ollama query gemma3 "¿Cuál es la capital de Japón?"

Kana iwe uchida kutora mukana we multimodal mabasa, unogona zvakare kushandisa mifananidzo mumibvunzo yako:

ollama query gemma3 --image "ruta-de-la-imagen.jpg"

Chii chaunoda kuti uiite zvakanaka? Kunyangwe Ollama asingamanikidzi zvakaomesesa zvinodikanwa, mhando hombe (se27B) inoda inokwana 32GB ye RAM. Ne 16GB unogona kushanda pasina matambudziko ne7B modhi, uye kunyangwe kushandiswa kweGPU kusingasungirwe, kunobatsira zvakanyanya mukumhanya.

Sarudzo yechipiri: Shandisa LM Studio

LM Studio ndechimwe chishandiso chemahara chinokutendera kuti uise uye umhanye mhando dzeLLM munharaunda kubva kune graphical interface.. Iyo inoenderana neWindows, macOS, uye Linux, uye mukana wayo mukuru ndeyekuti haidi ruzivo rwehunyanzvi kuti ushande.

Mirayiridzo:

1. Dhawunirodha LM Studio kubva kune yayo yepamutemo webhusaiti: lmstudio.ai.
2. Isa uye umhanye.
3. Dzvanya pachiratidzo chegirazi chekukudza chinoti "Ziva."
4. Nyora "Gemma 3" muinjini yekutsvaga kuti uone mamodheru aripo.

Usati waisa, tarisa kana modhi yacho ichienderana nemidziyo yako. Kana iwe ukaona yambiro "Ingangove yakakura kwazvo kumuchina uyu," unogona kuiisa, asi kuita kwakakwana hakuna kuvimbiswa.

Kana imwe modhi inoenderana yatorwa pasi:

• Dzvanya "Load Model" kuti uiise.
• Kana kuvhura chat itsva uye sarudza modhi kubva pane yekudonha-pasi menyu.

Chinhu chakanakisa nezve LM Studio ndechekuti inoita senge yemuno ChatGPT, isina online uye mumutauro wako.. Iwe unogona kugadzira akawanda chats uye chengetedza hurukuro dzako kana uchida. Pamusoro pezvo, kana iwe ukagonesa iyo "Local Server" sarudzo, unogona kuibatanidza nePython yako yekushandisa uchishandisa OpenAI-inoenderana API.

Sarudzo 3: Shandisa Google AI Studio (pamhepo)

Kana iwe usingakwanise kana kusada kuisa chero chinhu, unogona kushandisa Gemma 3 zvakananga kubva mugore neGoogle AI Studio.. Hapana kuisirwa kunodiwa, asi kubatana kweInternet neGoogle account inodiwa.

Iwe unongofanira kuenda aistudio.google.com uye sarudza "Gemma 3" kubva pane rondedzero yemamodheru. Kubva panguva iyoyo zvichienda mberi, unogona kutanga kutaura nemuenzaniso sekunge ishanduro yepamusoro yeBard kana ChatGPT, kusanganisira kuisa mufananidzo.

Kuiswa kwegore neNodeShift (kusarudza)

Kune avo vari kutsvaga simba rakawanda kana kuendesa modhi nehunyanzvi, pane sarudzo yekushandisa makore masevhisi seNodeShift. Nadzo, unogona kuhaya michina ine maGPU ane simba uye kugadzirisa yako yakanakira nharaunda kumhanya Gemma 3 pasina miganhu.

Matanho ekutanga muNodeShift:

1. Gadzira account ku app.nodeshift.com.
2. Tanga tsika yeGPU Node (semuenzaniso ine 2x RTX 4090).
3. Sarudza mufananidzo wakafanorongwa neUbuntu + Nvidia CUDA kana Jupyter Notebook, zvichienderana nekuti unenge uchishandisa Ollama kana Transformers.
4. Batanidza neSSH uye isa modhi kubva pamutsetse wekuraira.

Iyi mhando yekuisa inokupa iwe kuwana kune nyanzvi zvigadziriso, yakanakira mamodeli ekudzidzisa, kuongorora maitiro, nezvimwewo. Kunyange zvisingakoshi kune vashandisi vemba, inobatsira kune avo vanoda kuedza zvakadzama kana kuvaka maapplication paLLMs dzemberi.

Zvido zveSistimu uye kurudziro yehunyanzvi

Haasi ese maGemma 3 modhi anomhanya pane chero PC. Pazasi tinokusiira iwe general reference zvinoenderana nerudzi rwemodhi.

• Para mhando 1B kusvika 7B: zvishoma 8 GB ye RAM. Ivo vanoshanda pane ingangoita chero yemazuva ano PC, kunyangwe isina GPU.
• Para 13B mhando: zvinokurudzirwa 16GB kusvika 24GB RAM.
• Para 27B mhando: zvinodiwa ingangoita 32 GB ye RAM uye zvichida GPU yakatsaurirwa.

Kuva ne RAM yakawanda kunomhanyisa kushanda uye kunodzivirira kukanganisa nekuda kwekushaya ndangariro. Nepo Ollama neLM Studio vachiedza kushandisa zviwanikwa nemazvo, zvinoenderana nezvakawanda pane yako hardware. Pamusoro pezvo, kumhanya kwekupindura kunovandudza zvakanyanya kana GPU ikashandiswa pachinzvimbo cheCPU.

Kuisa Gemma 3 pa Windows 11 zviri nyore kupfuura zvinoita.. Izvo hazvina basa kana iwe ukafunga kushandisa Ollama nekureruka kwayo, LM Studio kune yayo graphical interface, kana Google AI Studio kuitamba yakachengeteka mugore. Chinhu chakakosha ndechekuti imwe neimwe nzira inoenderana neakasiyana mazinga echiitiko uye tekinoroji kugona. Zvino zvawava kuziva ese sarudzo uye izvo zvaunoda kuti utange, unogona kutanga kuyedza neiyi inoshamisa yenzvimbo yekugadzira njere nhasi.