Yakha Eyakho Imodeli ye-AI nge-Brewery: Isifundo Sokuyilungisa Isinyathelo Ngesinyathelo

Yini i-Brewery?

I-Brewery iyi-ejenti eqondisayo yokulungisa kahle amamodeli olimi nawezithombe. Uchaza ukuthi yini ofuna ukuyakha engxoxweni nomqondisi we-AI omkhethile, obizwa ngokuthi umqambi wobhiya. I-Brewery ibe isikusiza ichaze umgomo, ikhethe imodeli eyisisekelo, ilungise idatha, ikhethe ama-hyperparameter aphephile, iqhube ukuqeqeshwa, ihlole umphumela, bese ishicilela imodeli eqediwe.

Umqambi wobhiya angaba uClaude, imodeli ehambisana ne-OpenAI, i-OpenRouter, noma imodeli yasendaweni esebenza nge-Ollama noma i-LM Studio. I-Brewery ilungisa izincazelo namathuluzi ayo ngokwezinga lolwazi lwakho: Osaqalayo, Umsebenzisi wokuzilibazisa, Umakhi, noma Uchwepheshe.

Ukuqeqeshwa kungaqhutshwa ku-GPU yakho ye-NVIDIA noma ye-AMD, noma kuseva eqashiwe ofinyelela kuyo nge-SSH. Ikhompyutha ephathekayo esebenzisa i-Brewery idinga kuphela ukusebenza njengendawo yokulawula lapho ukuqeqeshwa kwenzeka kwenye indawo.

Izici Eziyinhloko

• Lungisa kahle amamodeli ombhalo asekelwayo usebenzisa i-LoRA, i-QLoRA, noma ukulungisa kahle okuphelele.
• Qeqesha amamodeli ombhalo ngokusebenzisa ukuqeqeshwa kwangaphambili okuqhubekayo, ukulungisa kahle okugadiwe, nokulungisa okusekelwe kokuncanyelwayo.
• Qeqesha ama-adapter e-Qwen-Image 2.1 LoRA ezitayela, kubalingiswa, nasezintweni.
• Ngenisa idatha ku-Hugging Face, kumafayela asendaweni, noma ezibonelweni ezidalwe nomqambi wobhiya.
• Ngokuzikhethela, dala idatha yokwenziwa noma amapheya okuncanyelwayo ngemva kokubuyekeza imigomo yomhlinzeki.
• Sebenzisa i-Empero Trace Format, noma i-ETF, ezingxoxweni, izingcingo zamathuluzi, ekucabangeni, kumadokhumenti e-RAG, kokuncanyelwayo, embhalweni ongahluziwe, ekuqedweni, nasezithombeni.
• Linganisa inkumbulo, isikhathi, nezindleko ngaphambi kokuqala ukuqeqeshwa.
• Qhuba ukuqeqeshwa endaweni noma ku-Runpod, ku-Vast.ai, noma kwenye iseva ye-SSH.
• Thumela ama-adapter noma amamodeli ahlanganisiwe ngekhadi lemodeli elakhiwe namafayela elayisensi.
• Layisha imiphumela ku-Hugging Face, ibe yimfihlo ngokuzenzakalela.

Faka i-Brewery

I-Brewery ifakwa kusuka endaweni yokugcina ye-GitHub nge-pip:

pip install "brewery-ai @ git+https://github.com/empero-org/brewery-ai"

Qalisa uhlelo ngokuthi:

brewery

Ekuqaliseni kokuqala, i-Brewery ibuza ukuthi iyiphi i-AI okufanele ikuqondise futhi icele ukhiye wayo we-API. Izifakazelo zigcinwa endaweni kumafayela afundwa nguwe kuphela. Ukusetha kuphinde kubuze izinga lolwazi lwakho negama lephrojekthi ngaphambi kokudala ifolda yephrojekthi.

Faka Ukusekelwa Kokuqeqeshwa kwe-GPU Yasendaweni

Uma uhlela ukuqeqesha ku-GPU ye-NVIDIA, faka izengezo zokuqeqesha. Lo myalo uphinde ufake i-PyTorch ehambisana ne-CUDA:

pip install "brewery-ai[train] @ git+https://github.com/empero-org/brewery-ai"

Kuma-GPU e-AMD, qala ufake inguqulo ye-PyTorch ye-ROCm efanele, bese ufaka i-Brewery nezengezo zokuqeqesha. Ku-Windows, gcina PIP_CONSTRAINT=constraints-windows-rocm.txt isethiwe ukuze i-pip ishiye ukufakwa kwe-PyTorch ye-AMD endaweni.

I-Apple Silicon namasistimu asebenzisa i-CPU kuphela afanele ukuqaliswa kokuhlola okuncane, kanti umqambi wobhiya angakuqondisa ekuqasheni i-GPU ukuze usebenze ngamamodeli amakhulu.

Qala Ukupheka Kwakho Kokuqala Okuqondiswayo

Ngemva kokuqalisa i-brewery, chaza ngolimi oluvamile imodeli ofuna ukuyakha. Isibonelo:

a chatbot that explains chemistry like a friendly pirate

Umqambi wobhiya uguqula lo mbono ube uhlelo lokuqeqesha olucacile. Ungacabangela ukuziphatha okufunwayo, ihardware etholakalayo, ubukhulu bemodeli, indlela yokuqeqesha, izidingo zedatha, izindleko ezilinganisiwe, nemikhawulo yelayisensi. Uma ikhompyutha yamanje ingenayo i-GPU efanele, ingancoma i-GPU eqashiwe engu-24 GB futhi ilinganise izindleko zokusebenza ezisondele.

Zonke izenzo ezisebenzisa imali, ezifaka isofthiwe, ezithumela idatha kwenye indawo, noma ezishicilela imodeli zidinga ukuqinisekiswa ngaphakathi kwe-Brewery. Umqondisi we-AI akakwazi ukweqa lezo zicelo, futhi amaphasiwedi, okhiye be-API, namathokheni awadluliswa engxoxweni.

Landela Inqubo Ye-Brewery

1. Chaza Umgomo

Chaza ukuthi imodeli kufanele yenzeni, yisiphi isitayela okufanele isisebenzise, yiziphi izihloko okufanele izazi, noma yimaphi amathuluzi namafomethi okufanele iwasekele. Qhubeka uphendula imibuzo yomqambi wobhiya kuze kube yilapho umgomo ucacile ngokwanele ukuba uqeqeshwe futhi uhlolwe.

2. Khetha Amandla Ekhompyutha

Sebenzisa brewery hardware ukuze uhlole isistimu yamanje:

brewery hardware

Ngaphambi kokuqeqeshwa kwasendaweni, i-Brewery yenza umsebenzi wangempela we-GPU ukuze izinkinga zomshayeli noma ze-PyTorch zivele ngaphambi kokuqala komsebenzi omude. Ekuqeqeshweni okukude, i-Brewery ikusiza ulungise iseva ye-SSH, ulayishe idatha edingekayo, bese uqhuba umsebenzi uzimele ukuze uqhubeke noma uvale i-terminal yakho.

3. Khetha Imodeli Eyisisekelo

I-Brewery isekela imindeni yombhalo ehlanganisa i-Qwen3, i-Qwen3.5, izakhiwo ezihlobene ne-Qwen, i-Qwen3.5 MoE, i-Llama 3.1 kuya ku-3.3, i-Gemma 3, ne-Gemma 4. Ekuqeqeshweni kwezithombe, isekela ama-adapter e-Qwen-Image 2.1 LoRA.

Sebenzisa isiphequluli samamodeli ukuze uhlole amamodeli ayisisekelo atholakalayo:

brewery models --modality text
brewery models --modality image
brewery models --all

Umqambi wobhiya ucabangela umgomo, ihardware, nezidingo zelayisensi. Uphinde ahlole ukufinyelela kumamodeli anemikhawulo lapho kudingeka khona.

4. Lungisa Idatha Yokuqeqesha

Ungasesha futhi uhlole kuqala amasethi edatha e-Hugging Face ngaphandle kokuwalanda, ungenise amafayela asendaweni, ubhale izibonelo nomqambi wobhiya, noma usebenzise isethi yedatha yezithombe ekhona. Idatha yokwenziwa iyazikhethela futhi yakhelwa ngokwemiyalelo yakho. I-Brewery ibonisa imigomo yomhlinzeki ngaphambi kokuyidala.

Isibonelo, ungacela umqambi wobhiya ukuthi abhale kabusha isethi yedatha ekhona ngezwi elithile, adale izingxoxo ngokwemiyalelo, noma akhiqize amapheya okuncanyelwayo e-DPO. Hlanza idatha engenisiwe ngokuthi:

brewery etf validate path/to/data.jsonl

Ngaphakathi kwephrojekthi, amasethi okuqeqesha agcinwa ngaphansi kwe-data/_sets/. I-Brewery ingasusa amarekhodi angenalutho, aphindaphindiwe, noma awonakele ngenkathi igcina ikhophi eyisipele.

5. Khetha Uhlelo Lokuqeqesha

Umsebenzi ngamunye unomgomo futhi ungaqala esigabeni sangaphambilini:

• SFT: ifundisa imodeli eyenzelwe ukulandela imiyalelo indlela yokuziphatha, isitayela, ifomethi, noma iphethini yokusebenzisa amathuluzi.
• CPT bese kuba yi-SFT: ifundisa imodeli eyisisekelo ngombhalo ongacutshungulwanga wesizinda, bese iyayiqeqesha ukuthi ixhumane ngendlela yengxoxo.
• SFT bese kuba yi-DPO: iqala ngokufundisa indlela yokuziphatha efiselekayo, bese icija izintandokazi isebenzisa amapheya ezimpendulo akhethiwe.

Izigaba ze-LoRA zihlanganiswa nezisindo zazo zokuqala ngaphambi kwesigaba esilandelayo, okwenza izindlela eziyizigaba ezixhunyiwe zinqwabelane kahle. Ithuluzi le-Brewery elithi propose_training_config ligcwalisa ama-hyperparameter ngokusekelwe emihlahlandleleni yamamodeli asekelwayo futhi lilinganisa iresiphi ngokwe-GPU etholakalayo.

6. Qalisa Ukuqeqesha

Qinisekisa iresiphi ephakanyisiwe ngemva kokubuyekeza inkumbulo, isikhathi, nezindleko ezilinganisiwe. Ukuqeqesha kuqhubeka ngaphandle kokuxhumeka emshinini okhethiwe, kuyilapho i-Brewery ibika inqubekelaphambili, ukulahleka, nesikhathi esilinganisiwe sokuqeda. Uma umsebenzi wehluleka, umphathi wotshwala unikeza incazelo ecacile ngolimi olujwayelekile kanye nesiphakamiso sokulungisa.

Ungahlola imisebenzi usebenzisa:

brewery status
brewery status JOB

Ukuvala i-terminal akuwumisi umsebenzi wokuqeqesha oqhubeka ngaphandle kokuxhumeka. Buyela kufolda yephrojekthi bese usebenzisa brewery ukuze uqhubeke neseshini.

7. Hlola Umphumela

Ngemva kokuqeqesha, hlola ama-prompt kwimodeli entsha bese uqhathanisa izimpendulo zayo nemodeli eyisisekelo. Ekulungiseni izintandokazi, i-Brewery ingakhiqiza abakhethiwe ababili nge-prompt ngayinye. Ungakhetha impendulo engcono ku-terminal, noma ucele i-AI eqondisayo ukuthi ihlole nge-rubric bese uhlola ngokwakho ezinye zezinqumo zayo.

Imiyalo yeseshini ewusizo ihlanganisa:

• /status ibonisa iphrojekthi yamanje nesimo somsebenzi.
• /jobs ibala imisebenzi yokuqeqesha.
• /watch ilandelela umsebenzi osebenzayo.
• /level ishintsha izinga lencazelo.
• /usage ibonisa ulwazi lokusetshenziswa.
• /help ibala imiyalo etholakalayo.

8. Khipha Futhi Ushicilele

Uma umphumela usulungile, i-Brewery ingakhipha i-adapter noma imodeli ehlanganisiwe, ikhiqize ikhadi lemodeli, ifake amafayela elayisensi afanele, bese ilayisha imodeli ku-Hugging Face. Amakhadi amamodeli angafaka indlela yokuqeqesha, imithombo yedatha, amalayisensi, amakhredithi, namasampula angaphambi nangemva.

Ukulayisha kuyimfihlo ngokuzenzakalela. Buyekeza ikhadi lemodeli nezimfuneko zelayisensi ngaphambi kokwenza indawo yokugcina ibe sesidlangalaleni. Uma usebenzisa ingqalasizinda eqashwayo, khumbula ukumisa noma ukubhubhisa iseva: ama-pod nama-instance amisiwe asengakhokhisa isitoreji.

Qeqesha ama-LoRA Ezithombe Ngama-Checkpoint Wokubuka Kuqala

Emisebenzini ye-Qwen-Image 2.1 LoRA, i-Brewery ikhiqiza izithombe zokubuka kuqala ngesikhathi sokuqeqesha. Idala isethi eyisisekelo esinyathelweni esingu-zero kanye namanye amasethi kuma-checkpoint, isebenzisa ama-seed afanayo ukuze kube lula ukuqhathanisa okubonakalayo.

Amafayela okubuka kuqala alandwa ezindleleni ezifana ne-runs/<job>/samples/step_NNNNNN/. Isethi ngayinye yokubuka kuqala ine-checkpoint ehambisanayo. Uma i-checkpoint yangaphambilini ibukeka ingcono kunomphumela wokugcina, cela umphathi wotshwala ukuthi apakishe leyo checkpoint.

Sebenzisa /gallery phakathi neseshini ukuze uqale igalari yasendaweni eseceleni nesehlangothini. Iseva kukhompyutha yakho ekhelini lendawo elifana ne-http://127.0.0.1:8765/<random-token>/ futhi ayifinyeleleki ngaphandle. Igalari iyazibuyekeza lapho kufika ukubuka kuqala okusha, ikuvumela ukuthi ukhulise izithombe, futhi isekela ukuqhathanisa nemodeli eyisisekelo.

Amathiphu Athuthukile

Khetha Ngokucophelela I-AI Eqondisayo

I-Claude iyinketho enconyiwe ukuze uthole ukuqondiswa okuqinile, kuyilapho abahlinzeki abahambisana ne-OpenAI kanye ne-OpenRouter benikeza ukukhetha okubanzi kwamamodeli. I-Ollama ne-LM Studio zinikeza ukusebenza kwasendaweni; i-Brewery incoma imodeli yasendaweni engu-14B noma enkulu enokubiza amathuluzi ukuze uthole imiphumela emihle. Amamodeli angenakho ukubiza amathuluzi okwendabuko asengaqondiswa ngephrothokholi yombhalo ezenzakalelayo.

Shintsha imodeli eqondisayo noma nini usebenzisa:

brewery setup

Sebenzisa i-ETF Kumapayipi Edatha Avumelanayo

I-ETF inikeza ifomethi eyodwa ye-JSONL yezinto zokuqeqesha ezihlukahlukene, okuhlanganisa izingxoxo, ama-system prompt, ama-schema namakholi wamathuluzi, ukucabanga, imibhalo ye-RAG, amapheya ezintandokazi, umbhalo ongacutshungulwanga, ama-completion, nezithombe. Lokhu kwenza kube lula ukuqinisekisa nokudlulisa idatha phakathi kwezigaba zephrojekthi. Bheka imibhalo ye-ETF ukuze uthole imininingwane yefomethi.

Lungisa Ukuziphatha Kwe-terminal

Kuma-terminal anendawo elinganiselwe, sebenzisa okuguquguqukayo kwemvelo ukuze wenze isixhumi esibonakalayo sibe lula noma ulawule ukukhiqizwa kwedatha yokwenziwa:

BREWERY_AI_PLAIN_INPUT=1
BREWERY_AI_SIMPLE_PROMPTS=1
BREWERY_AI_THINKING=off
BREWERY_AI_SYNTH_WORKERS=4

BREWERY_AI_SYNTH_WORKERS yamukela amanani asuka ku-1 kuya ku-16. Ukuboniswa kokucabanga kungasethwa ku-on, full, noma ku-off.

Qonda Ukwakheka Kwephrojekthi

Iphrojekthi ye-Brewery igcina izinqumo, amasethi edatha, imisebenzi, imiphumela, nomlando wezingxoxo ndawonye:

my-project/
  brewery.yaml
  data/
  data/_sets/<name>/
  runs/<job_id>/
  .brewery/

Izindlela zigcinwa zihlobene nephrojekthi kusetshenziswa ama-slash aya phambili, ngakho iphrojekthi ingahanjiswa phakathi kwamakhompyutha nezinhlelo zokusebenza. Izilungiselelo zomsebenzisi zigcinwa ku-~/.config/brewery-ai/ noma, ku-Windows, ku-%APPDATA%\brewery-ai\.

Isiphetho

I-Brewery iguqula ukulungisa kahle imodeli kube ukuhamba komsebenzi okuqondisiwe futhi okuhlolwayo. Faka uhlelo lokusebenza lokulawula, khetha umphathi wotshwala onekhono, chaza umgomo wakho, lungisa futhi uqinisekise idatha, uvume iresiphi eqaphela ihardware, bese uvumela i-Brewery ilawule ukuqeqesha, ukuhlola, ukukhipha, nokushicilela. Qala ngephrojekthi encane ye-SFT, bese uhlola izindlela ezixhunyiwe ze-CPT, SFT, ne-DPO noma ama-LoRA ezithombe njengoba izidingo zakho zikhula.