Momwe mungakhalire Gemma 3 LLM pa Windows 11 sitepe ndi sitepe

Zilankhulo zotseguka zoyambira zidasintha modumphadumpha, ndi Masiku ano ndizotheka kusangalala nawo mwachindunji kuchokera pakompyuta yanu popanda kutengera ntchito zamtambo.. Chimodzi mwazosangalatsa kwambiri pakadali pano ndi Gemma 3, LLM yatsopano ya Google yozikidwa paukadaulo wa Gemini, womwe umadziwika bwino chifukwa cha kuthekera kwake polemba zolemba ndi zithunzi, komanso pazenera lake lalikulu la ma tokeni mpaka 128k m'mitundu yake yapamwamba. Kuti mudziwe zambiri za kumasulidwa uku, mukhoza kupita ku nkhani yathu chiwonetsero cha Gemma 3.

Ngati mukugwiritsa ntchito Windows 11 ndikuyang'ana kukhazikitsa Gemma 3 poyesera kapena kupanga kwanuko, mwafika pamalo oyenera. Tiyeni tidutse mwatsatanetsatane njira zonse zomwe mungakhazikitsire pakompyuta yanu, kuphatikiza zosankha zomwe mwalimbikitsa kwambiri monga Ollama, LM Studio, komanso njira ina yozikidwa pamtambo ndi Google AI Studio. Kuonjezera apo, tidzakambirana zofunikira zamakono, ubwino wa njira iliyonse ndi Momwe mungapangire mwayi wanzeru zamphamvu zopangazi.

Kodi Gemma 3 ndi chiyani ndipo chifukwa chiyani muyiyikire?

Gemma 3 ndi m'badwo wachitatu wamitundu ya LLM yotulutsidwa ndi Google pansi pa layisensi yotsegula.. Mosiyana ndi mayankho am'mbuyomu monga Llama kapena Mistral, imapereka chithandizo chachindunji pakuyika zithunzi, mawonekedwe okulirapo, ndikuthandizira zilankhulo zopitilira 140. Kutengera zosowa zanu, mutha kusankha kuchokera kumitundu ingapo yachitsanzo kuyambira 1B mpaka 27B magawo:

• Gemma 3:1B: Mtundu wopepuka wokwanira pantchito zoyambira komanso malo opanda zida.
• Gemma 3:4B: Kulinganiza magwiridwe antchito ndi magwiridwe antchito apakatikati.
• Gemma 3:12B: Akulimbikitsidwa kusanthula kovutirapo, kukonza mapulogalamu ndi zinenero zambiri.
• Gemma 3:27B: Njira yamphamvu kwambiri, yopangidwira kugwiritsa ntchito mozama, ma multimodal okhala ndi mphamvu zambiri.

Kuthekera koyendetsa mtundu wapamwamba kwambiri kuchokera pa PC yanu Imasintha malamulo a masewerawa ponena zachinsinsi, liwiro la kuyankha, ndi kudalira anthu ena. Simufunikanso kulipira zolembetsa pamwezi kapena kusiya deta yanu. Zomwe zimafunika ndikukonzekera pang'ono ndi kufuna kuphunzira. Ngati mukufuna kudziwa zambiri za ubwino wa zitsanzozi, onani nkhani yathu Openweight AI zitsanzo.

Njira 1: Kuyika ndi Ollama

Ollama mwina ndiye njira yosavuta yoyendetsera LLM ngati Gemma 3 kuchokera Windows 11. Mawonekedwe ake otengera ma terminal amakulolani kuti muyike ndikuyendetsa zitsanzo ndi mzere wosavuta wamalamulo. Kuphatikiza apo, imagwirizana ndi macOS, Linux, ndi Windows, ndikupangitsa kuti ikhale yosavuta kugwiritsa ntchito m'malo osiyanasiyana.

Njira zoyika Ollama ndikuyendetsa Gemma 3:

1. Pezani patsamba lovomerezeka: olama.com.
2. Koperani okhazikitsa kwa Windows ndikuyendetsa ngati pulogalamu ina iliyonse.
3. Tsegulani Command Prompt (CMD) kapena PowerShell ndikutsimikizira kukhazikitsa ndi:

ollama --version

Zonse zikayenda bwino, mutha kutsitsa ma tempulo aliwonse a Gemma 3 omwe alipo. Ingoyendetsani limodzi mwamalamulo awa kutengera template yomwe mukufuna:

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

Kamodzi dawunilodi, inu mosavuta kuyamba chitsanzo.. Kuti muchite izi, yesani:

ollama init gemma3

Kuyambira nthawi imeneyo, mutha kuyamba kucheza ndi LLM mwa:

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

Ngati mukufuna kugwiritsa ntchito ma multimodal, mutha kugwiritsanso ntchito zithunzi pamafunso anu:

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

Kodi muyenera kuchita chiyani kuti mugwire bwino? Ngakhale Ollama samakhazikitsa zofunikira zochepa, zitsanzo zazikulu (monga 27B) zimafuna osachepera 32GB ya RAM. Ndi 16GB mungathe kugwira ntchito popanda mavuto ndi chitsanzo cha 7B, ndipo ngakhale kugwiritsa ntchito GPU sikuli kovomerezeka, kumathandiza kwambiri mofulumira.

Njira 2: Gwiritsani ntchito LM Studio

LM Studio ndi chida china chaulere chomwe chimakupatsani mwayi woyika ndikuyendetsa mitundu ya LLM kwanuko kuchokera pazithunzi.. Ndiwogwirizana ndi Windows, macOS, ndi Linux, ndipo mwayi wake waukulu ndikuti safuna chidziwitso chaukadaulo kuti ugwire ntchito.

Malangizo:

1. Tsitsani LM Studio patsamba lake lovomerezeka: lmstudio.ai.
2. Ikani ndikuyendetsa.
3. Dinani pa chithunzi cha galasi lokulitsa lomwe likuti "Discover."
4. Lembani "Gemma 3" mu injini yosakira kuti muwone mitundu yomwe ilipo.

Musanayike, fufuzani ngati chitsanzocho chikugwirizana ndi zipangizo zanu. Ngati muwona chenjezo lakuti "Mwachidziwitso chachikulu kwambiri pamakinawa," mutha kuyiyikabe, koma kuchita bwino sikunatsimikizidwe.

Mtundu wogwirizana ukatsitsidwa:

• Dinani "Load Model" kuti muyitse.
• Kapena tsegulani macheza atsopano ndikusankha chitsanzo kuchokera pa menyu yotsitsa.

Zabwino kwambiri za LM Studio ndikuti imachita ngati ChatGPT yakomweko, osagwiritsa ntchito intaneti komanso m'chilankhulo chanu.. Mutha kupanga macheza angapo ndikusunga zokambirana zanu ngati mukufuna. Kuphatikiza apo, ngati muthandizira njira ya "Local Server", mutha kuyiphatikiza ndi mapulogalamu anu a Python pogwiritsa ntchito OpenAI-compatible API.

Njira 3: Gwiritsani ntchito Google AI Studio (pa intaneti)

Ngati simungathe kapena simukufuna kukhazikitsa chilichonse, mutha kugwiritsa ntchito Gemma 3 kuchokera pamtambo ndi Google AI Studio.. Palibe kukhazikitsa komwe kumafunikira, koma intaneti ndi akaunti ya Google ndizofunikira.

Muyenera kungopita aistudio.google.com ndikusankha "Gemma 3" pamndandanda wamitundu. Kuyambira nthawi imeneyo, mukhoza kuyamba kucheza ndi chitsanzocho ngati kuti ndi mtundu wapamwamba wa Bard kapena ChatGPT, kuphatikizapo kujambula zithunzi.

Kuyika kwamtambo ndi NodeShift (posankha)

Kwa iwo omwe akufunafuna mphamvu zambiri kapena kutumiza chitsanzo mwaukadaulo, pali mwayi wogwiritsa ntchito mautumiki amtambo monga NodeShift. Ndi iwo, mutha kubwereka makina okhala ndi ma GPU amphamvu ndikusintha malo anu abwino kuti muyendetse Gemma 3 popanda malire.

Njira zoyambira mu NodeShift:

1. Pangani akaunti pa app.nodeshift.com.
2. Yambitsani chizolowezi cha GPU Node (mwachitsanzo ndi 2x RTX 4090).
3. Sankhani chithunzi chokonzedweratu ndi Ubuntu + Nvidia CUDA kapena Jupyter Notebook, kutengera ngati mukugwiritsa ntchito Ollama kapena Transformers.
4. Lumikizani kudzera pa SSH ndikuyika chitsanzo kuchokera pamzere wamalamulo.

Kuyika kwamtunduwu kumakupatsani mwayi wofikira masanjidwe aukadaulo, yabwino kwa zitsanzo zophunzitsira, kuyesa ntchito, ndi zina zotero. Ngakhale sizofunikira kwa ogwiritsa ntchito kunyumba, ndizothandiza kwa iwo omwe akufuna kuyesa mozama kapena kupanga mapulogalamu pa LLMs apamwamba.

Zofunikira pamakina ndi malingaliro aukadaulo

Si mitundu yonse ya Gemma 3 yomwe idzayendetse pa PC iliyonse. Pansipa tikusiyirani mbiri yofananira ndi mtundu wamitundu:

• Para mitundu 1B mpaka 7B: osachepera 8 GB ya RAM. Amagwira ntchito pafupifupi PC iliyonse yamakono, ngakhale opanda GPU.
• Para Zithunzi za 13B: amalimbikitsidwa 16GB mpaka 24GB RAM.
• Para Zithunzi za 27B: zofunika osachepera 32 GB ya RAM ndipo makamaka GPU yodzipereka.

Kukhala ndi RAM yochulukirapo kumathandizira kugwira ntchito ndikuletsa zolakwika chifukwa chosowa kukumbukira. Ngakhale Ollama ndi LM Studio amayesa kugwiritsa ntchito zinthu moyenera, zimatengera kwambiri zida zanu. Kuphatikiza apo, liwiro loyankhira limayenda bwino kwambiri ngati GPU imagwiritsidwa ntchito m'malo mwa CPU.

Kuyika Gemma 3 Windows 11 ndikosavuta kuposa momwe zimawonekera.. Zilibe kanthu ngati mutasankha kugwiritsa ntchito Ollama chifukwa chophweka, LM Studio pazithunzi zake, kapena Google AI Studio kuti muzisewera bwino pamtambo. Chofunikira ndichakuti njira iliyonse igwirizane ndi magawo osiyanasiyana odziwa komanso luso laukadaulo. Tsopano popeza mukudziwa zonse zomwe mungachite komanso zomwe muyenera kuchita kuti muyambe, mutha kuyamba kuyesa luntha lochita kupanga la m'deralo lero.