- Ukuhlanganiswa okugcwele kwe-Tensor ML SDK ne-LiteRT ukuze kuthuthukiswe ukuthunyelwa kwamamodeli e-AI kumadivayisi e-Pixel.
- Ukufinyelela Engadini Yamamodeli enamamodeli angaphezu kwe-100 alungiselelwe i-Google Tensor TPU.
- Ukusekelwa okuthuthukisiwe kwamamodeli olimi olunezindlela eziningi kanye nombono ngokusebenzisa ukwakheka kwe-Gemma 4.
- Ikhono lokwenza iziphetho zesivinini esiphezulu kanye nobumfihlo ngenxa yehadiwe ye-Tensor G5.
Uma uthanda ukuthuthukiswa kohlelo lokusebenza kanye nobuhlakani bokwenziwa, cishe uke waphawula ukuthi umdlalo uyashintsha. I-Google inqume ukuvula amandla ehadiwe yayo enamandla kakhulu nge- Tensor ML SDK , ithuluzi lokuthuthukisa elivumela abathuthukisi ukuthi basebenzise ngokugcwele iyunithi yokucubungula i-tensor (TPU) yamadivayisi e-Pixel, okwenza i-AI isebenze ngqo kudivayisi, kunokuthembela efwini.
Into ejabulisa kakhulu ukuthi le SDK isidlule esigabeni sayo sokuhlola yaya ku- Beta . Lokhu kusho ukuthi akuseyona eyabambalwa abakhethiwe kuphela; noma yimuphi umhleli wezinhlelo manje angaqala ukudala okuhlangenwe nakho kwe-AI okusebenzisanayo, okuyimfihlo, futhi, ngaphezu kwakho konke, okusheshayo kakhulu ngokusebenzisa ukwakheka kwe-Tensor SoC ye-Google ukwenza imisebenzi eyayibonakala ingenakwenzeka ukuyenza endaweni.
Umsebenzi ohlangene ngenxa ye-LiteRT
Ukuze kuvinjelwe abathuthukisi ukuthi bangabambeki ekucushweni okuyinkimbinkimbi, i-Google ihlanganise i-Tensor SDK ne -LiteRT . Lolu hlaka lusebenza njengesendlalelo sokucashunwa, lususa isidingo sokubhekana nama-SDK athile kumthengisi noma abahlanganisi abayinkimbinkimbi futhi luhlinzeka nge-API elula nehambisanayo yokufaka amamodeli okufunda komshini emaphethelweni.
Le nqubo ihlukaniswe ngezigaba ezintathu ezibalulekile. Okokuqala ukuhlanganiswa kwemodeli , lapho ungaguqula khona amaphrojekthi akho asekelwe ku-PyTorch noma ku-TFLite abe yi-binary elungiselelwe usebenzisa i-LiteRT Torch. Okulandelayo kuza ukuthunyelwa , kusetshenziswa i-Play Feature Delivery kanye ne-AI Packs ukuze kusatshalaliswe kahle imitapo yolwazi ehlanganisiwe namamodeli ngaphakathi kohlelo lokusebenza.
Ekugcineni, sifika ekuqalisweni kwesinqumo. Ngenxa ye- LiteRT Runtime , ungenza imodeli yakho isebenze ku-TPU ngemigqa embalwa yekhodi. Okuhle kunakho konke, uhlelo luhlakaniphile: uma nganoma yisiphi isizathu i-TPU ingatholakali, ungalungiselela izindlela zokubuyela emuva ukuze uhambise ngokuzenzakalelayo umthwalo ku-CPU noma ku-GPU, uqinisekise ukuthi uhlelo lokusebenza alulokothi lubambeke.
Ingadi Yemodeli: Ikhathalogi yamathuba
Asikho isidingo sokuqala kusukela ekuqaleni, njengoba i-Beta SDK ifaka i- Model Garden emangalisayo . Lo mtapo wolwazi uqukethe amamodeli angaphezu kwe-100, kokubili i-ML yakudala kanye ne-AI ekhiqizayo, kufaka phakathi izinguqulo ze- Gemma 3 1B , kanye nenani elikhulu lamamodeli akhiwe ngaphambilini ongawalanda ngqo emphakathini we-Hugging Face ku-LiteRT.
Uma ufuna ukudala imisebenzi yombhalo, amamodeli amancane olimi afana ne -Function Gemma akuvumela ukuthi wenze izenzo zasendaweni ngaphakathi kohlelo lokusebenza, kuyilapho i-EmbeddingGemma ingeza amakhono athuthukile e-semantic. Ngasohlangothini olubonakalayo, i-SDK ikuvumela ukuthi usebenzise ukutholwa kwezinto kanye nokumapha okujulile , okubaluleke kakhulu ezinhlelweni zokusebenza zekhamera ezidinga ukusabela endaweni yomsebenzisi ngesikhathi sangempela.
Abakakhohlwa ngomsindo kanye nokufinyeleleka. Ukuqashelwa kwenkulumo kusukela ekuqaleni kuya ekugcineni sekungenzeka manje , kuqinisekisa ukulotshwa kwamagama ngamathuluzi okubambezeleka aphansi kakhulu kanye nokuhumusha asebenza ungaxhunyiwe ku-inthanethi, kugcinwa ubumfihlo bedatha njengoba kungekho lutho oluphuma efonini.
Ukuthuthukiswa kobuchwepheshe kanye nokusekelwa kwehadiwe
Ukuze uthole okuningi kukho, kubalulekile ukwazi ukuthi iyiphi ihadiwe esekelwayo. Njengamanje, i-ecosystem igxile emndenini wePixel 10 , okuhlanganisa amamodeli e-Pro, Pro XL, kanye ne-Pro Fold, wonke ahlonyiswe nge- Google Tensor G5 SoC . Ukuqinisekisa ukuthi i-AI isebenza kahle, i-Google incoma ukusebenzisa amafulegi athile okwenza ngcono ngesikhathi senqubo yokwakha.
Isibonelo, ekugelezeni kwe-LiteRT Python, kuvamile kakhulu ukusebenzisa ifulegi elithi `google_tensor_truncation_type="half"` ukuze kulungiswe ukusebenza kanye nokusetshenziswa kwezinsiza. Kumamodeli ezilimi ezinkulu (ama-LLM), ukuthumela ngaphandle kudinga amapharamitha anemininingwane, njengokulungiselela ` i-quantization_recipe` kanye nokuvumela ukusekelwa kwamamodeli amakhulu ngesichazamazwi sokucushwa kwe-AOT.
Kubalulekile ukusho ukuthi, yize i-NNAPI yayikhona ngaphambilini, ibilokhu ingasebenzi kusukela ku-Android 15. Indlela yamanje iwukudlulisa konke kubamele be-LiteRT, lapho ukusekelwa kwe-Pixel TPU sekuyinto ebalulekile ekushintsheni ukuqaliswa okudala kanye nokufeza ukusebenza kahle kwamandla okukhulu.
Uguquko lwe-Gemma 4 kanye ne-Multimodal AI
Ake sixoxe ngezinto zakamuva: ukufika kwe -Gemma 4 12B . Ngokungafani namanye amamodeli anamathisela i-encoder yesithombe kumodeli yolimi, i-Gemma 4 icubungula izithombe ngendlela yemvelo. Lo mklamo olula awugcini nje ngokunciphisa ukusetshenziswa kwe-VRAM kodwa futhi uvumela ukucabanga okubushelelezi nokuvumelanayo.
Njengoba inefasitela lomongo lamathokheni angu-256K , le modeli ingaphatha izingxoxo ezinde ngezithombe eziningi ngaphandle kokulahlekelwa umkhondo. Ngaphezu kwalokho, njengoba isatshalaliswa ngaphansi kwelayisensi ye -Apache 2.0 , iguquguquka kakhulu ekusetshenzisweni kwezentengiselwano nasekusatshalalisweni kabusha, ivumela i-AI ye-multimodal ukuthi isebenze kuma-laptops esimanje kusetshenziswa i-quantization engu-4- noma engu-8-bit.
Umgomo we-Google ucacile: ufuna abathuthukisi basebenzise amamodeli ayo okulinganisa isisindo ukuze balawule uhlelo lwe-ecosystem. Ngokwenza i-AI ibe yendawo futhi ibe namandla , kunciphisa ukuthembela kuma-API angaphandle futhi kwakha umphakathi ohlangene ozungeze ihadiwe ye-Tensor kanye nesofthiwe ye-LiteRT.
Uhlelo lokuthuthukiswa kwePixel luthathe igxathu elibalulekile ngokuhlanganisa ihadiwe ye -Tensor G5 nokuguquguquka kwe-LiteRT kanye namandla amamodeli afana ne-Gemma 4. Ngenxa yokushintshela ku-Beta ye-SDK kanye nokutholakala kwekhathalogi enkulu yamamodeli alungiselelwe kusengaphambili, ukudala izinhlelo zokusebenza ezicubungula umbono, inkulumo, kanye nolimi ngasese nangokushesha kakhulu manje sekuyinto efinyelelekayo kunoma yimuphi umhleli wezinhlelo.
