La Ubukrelekrele bokwenziwa obusebenza kwindawo yakho kwiWindows 11 Ayisekho kwiilabhoratri okanye kwiinkampani ezinkulu kuphela. Namhlanje ungaqhuba iimodeli eziphambili ngqo kwiPC yakho, ngaphandle kokuxhomekeka kwilifu, usebenzise iCPU, iGPU, kwane-NPU ukuba umatshini wakho usemgangathweni ophezulu. Kwaye isitshixo sokufezekisa oku lula... I-ONNX Runtime idityaniswe neWindows ML kunye ne-.NET.
Kule migca ilandelayo siza kubona, ngokuzolileyo kodwa ngaphandle kokubetha ecaleni kwehlathi, Indlela yokusebenzisa i-AI yendawo kunye ne-ONNX Runtime kwiWindows 11Yintoni i-ONNX kunye nexesha layo lokusebenza, indlela ehambelana ngayo neWindows ML, into ekufuneka uyifake, indlela yokuseta imizekelo yehlabathi lokwenyani kwi-.NET/WinUI 3, indlela yokusebenzisa iDirectML kunye nokukhawuleziswa kwehardware, kunye neemeko ezisebenzayo onokuzigubungela (ukuhlelwa komfanekiso, intetho-ukuya-kumbhalo, iimodeli zolwimi lwasekuhlaleni, i-RAG, njl.njl.). Uza kubona ikhowudi ethile, kodwa ingqwalasela ikukuqonda kwakho ukuhamba okupheleleyo kwe Layisha, lungiselela, kwaye usebenzise iimodeli ze-ONNX kumatshini wakho.
Yintoni i-ONNX kwaye idlala yiphi indima i-ONNX Runtime kwiWindows 11?
Umgangatho I-ONNX (Vula iNeural Network Exchange) Yifomathi evulekileyo eyilelwe ukuchaza iimodeli zenethiwekhi ye-neural ngendlela esebenzisanayo. Ngamafutshane: ungenza qeqesha imodeli kwiPyTorch okanye kwiTensorFlowyithumele kwi-ONNX uze uyisebenzise kwiWindows, kwilifu, kwiwebhu, okanye naphi na apho ufuna khona, ngaphandle kokubhala yonke into kwakhona ukususela ekuqaleni.
Un Imodeli ye-ONNX Oku kuquka ulwakhiwo lwenethiwekhi (iileya, uqhagamshelo, iintlobo zokusebenza), ubunzima obuvela kuqeqesho, kunye nenkcazo yeengeniso kunye neziphumo. Ngenxa yoku, izixhobo ezahlukeneyo kunye namaxesha okusebenza zinokuyiqonda ngokuchanekileyo. ukuba balindele ntoni, indlela yokuyicubungula, kunye noko bakubuyiselayoUmzekelo, imodeli yokwahlulahlula umfanekiso ifumana umfanekiso osele usetyenzisiwe kwaye ibuyisela i-probability vector kwiklasi nganye.
Isekelwe kuloo fomathi iqhelekileyo Ixesha lokudlala le-ONNXeyinjini yexesha lokusebenza elungiselelweyo yokusebenzisa ezi modeli kumaqonga amaninzi. Eli xesha lokusebenza libonelela nge-API efanayo ye:
- Layisha iimodeli ze-ONNX kwidiski okanye kwimemori.
- Yenza iiseshoni zokuqikelela ngeendlela ezahlukeneyo zokwenza.
- Qhagamshela ababoneleli bokuphumeza (Ababoneleli boLawulo) be-CPU, i-GPU kunye ne-NPU.
- Yondla imodeli ngee-tensors zokufaka kwaye ufumane ii-tensioners zemveliso.
KwiWindows 11, i-ONNX Runtime idityaniswe ngokwendalo ngokusebenzisa Windows ML kunye nenkqubo yendalo I-Windows App SDK kunye ne-DirectMLOku kuvumela usetyenziso lwakho lwedesktop, nokuba lukwi-C#, WinUI 3, WPF, okanye usetyenziso lwewebhu oluneWebNN, ukuba lusebenzise ithuba Ukukhawuleziswa kwezixhobo zekhompyutha ngaphandle kobunzima babaqhubi kunye nee-SDK ezithile zomenzi.
Iingenelo zokusebenzisa i-AI apha ekhaya xa kuthelekiswa nelifu
Sebenza kunye Iimodeli ze-AI zasekhaya kwiWindows 11 Inezibonelelo ezininzi ezicacileyo ngaphezu kokuxhomekeka kuphela kwiinkonzo ezikude ezifana neChatGPT, Gemini, okanye ezinye:
Okokuqala, i ubumfihlo bedatha kunye nemfihlo Iphucuka kakhulu kuba ukucubungula kwenziwa kwisixhobo sakho. Xa usebenzisa imodeli ye-ONNX kwiPC yakho, amaxwebhu akho, isandi, imifanekiso, okanye iirekhodi zonyango aziphumi kwinethiwekhi yakho, nto leyo enciphisa kakhulu umngcipheko wokuvuza okanye ukusetyenziswa gwenxa kolwazi.
Kukwakho nenye into iindleko kunye nolawuloNgokungathembeli kwizicelo eziqhubekayo kwi-API yelifu ehlawulelwayo, uphepha amatyala angalindelekanga kwaye unokuhambisa izisombululo ezisebenza ngobuninzi kumbutho wonke ngaphandle kokuphinda-phinda iindleko zokubhalisa. Utyala imali kwi-hardware kube kanye kwaye ufumane okuninzi kuyo nge-ONNX Runtime kunye neWindows ML.
La Ukubambezeleka okuphantsi Eli lelinye iqhinga. Ukuba wenza iingqikelelo zombono wekhompyutha, abancedisi bomxholo, okanye ukuqaphela intetho, uhambo ngalunye oluya efini longeza ii-milliseconds. Ukusebenzisa imodeli kwindawo yakho kunciphisa ixesha lokuphendula kwaye kuphucula imvakalelo yokutyibilika, nto leyo ibalulekileyo kwizicelo ezisebenzisanayo okanye kwizixhobo ezisemaphethelweni.
Okokugqibela kodwa okungakuncinananga yi ukuzimela kuqhagamshelo lwe-IntanethiUkuba i-app yakho kufuneka isebenze kwinqwelo moya, kumzi-mveliso okude, okanye nakweyiphi na indawo enenethiwekhi encinci, ukuba neemodeli ze-ONNX ezifakwe ngaphakathi kwesicelo ngokwaso (umzekelo, ngaphakathi kwe-MSIX) kuqinisekisa ukuba i-AI iya kuqhubeka isebenza ngokungenasiphako, njengakwi I-PhotoPrism ene-AI yendawo.
IWindows ML, i-ONNX Runtime kunye neDirectML: indlela ezihambelana ngayo zonke
Kwinkqubo ye-Windows 11 yanamhlanje, IWindows ML isebenza njengomaleko we-middleware Idibanisa ulawulo lwe-CPU, i-GPU, kunye ne-NPU ukuze kuqondwe imodeli ye-AI. Umsebenzi wayo kukulungelelanisa izixhobo zehardware ezikhoyo ukuze ukusetyenziswa kwemodeli ye-ONNX kube nempumelelo kangangoko kunokwenzeka, ngaphandle kokuba ulwe neenkcukacha zetshiphusi nganye.
I-Windows ML yi idityaniswe ngokuqinileyo ne-ONNX RuntimeIsebenzisa kwakhona ii-API zayo, ixhomekeke kubaNiki bayo bokuSebenza (ii-EP), kwaye inika ukuhlanganiswa kunye nokulungiswa kwemodeli kuyo. IMicrosoft ikwaphatha usasaze kwaye ugcine zombini i-ONNX Runtime kunye nee-EP ezivela kubavelisi abahlukeneyoOku kwenza kube lula kakhulu ukupakishwa kwezicelo zakho kwaye kunciphisa ukuxhomekeka kwangaphandle.
Inxalenye ephambili kule phazili yile I-DirectMLUmaleko wokutsalwa kwemisebenzi okuvumela ukuba usebenzise iimodeli ze-ML ngokusebenzisa i-GPU (kunye ne-NPU kwiinkqubo ezihambelanayo) ngendlela edibeneyo. I-ONNX Runtime inikezela nge-EP ezinikeleyo ye-DirectML, ukuze ukwazi Yenza iiseshoni zokuqikelela ezisebenzisa i-CPU, i-GPU, okanye i-NPU ngokuzenzekelayo kuxhomekeke kwinto ekhoyo kwikhompyutha.
Abavelisi abakhulu bathanda I-AMD, i-Intel, i-NVIDIA kunye ne-Qualcomm Basebenzisana neMicrosoft ukubonelela ngee-EPs eziphuculiweyo: I-AMD idibanisa iRyzen AI, i-Intel idibanisa i-OpenVINO neWindows ML, i-NVIDIA ibonelela ngeTensorRT kwii-RTX GPUs, kwaye iQualcomm ilungisa iSnapdragon X NPU. Ngenxa yoku, iimodeli ezifanayo ze-ONNX zinokusebenzisa izixhobo ezahlukeneyo ngaphandle kokuba utshintshe ikhowudi yakho ephezulu; ukuba ufuna ukulungisa inkqubo, funda indlela yokwenza. lungiselela iiprofayili zokusebenza.
Konke oku kuza kupakishwe kwi I-Windows App SDK (ukususela kwinguqulelo 1.8.1 ukuya phambili) kwaye ixhaswa ngokusesikweni kwizixhobo ezine IWindows 11 24H2 okanye ngaphezuluOku kwenza inkqubo yokusebenza ibe liqonga eliqinileyo kakhulu lamava e-AI asekuhlaleni kwimveliso.
Izinto ezifunekayo ukuze usebenzise i-AI yendawo kunye ne-ONNX Runtime kwiWindows 11
Ukuqala Sebenzisa iimodeli ze-ONNX kwindawo yakho kwiWindows 11 Kwizicelo zakho ze-.NET, kuyacetyiswa ukuba uphonononge ezinye iimfuno ezisisiseko zendalo kunye nezixhobo:
Kwinqanaba lenkqubo, kuya kufuneka Windows 11 ihlaziyiwe (kungcono ukwakha i-22621 okanye ngaphezulu) kwaye, ukuba uza kusebenzisa ii-apps zanamhlanje, sebenza ne Iiprojekthi zeWindows App SDK kunye neWinUI 3 zipakishwe kwiMSIXOku kukunika ukufikelela kwi-Windows ML kunye ne-DirectML APIs, kunye neempawu zamva nje zokudibanisa i-AI.
Ngokuphathelele imeko-bume yophuhliso, ukhetho oluqhelekileyo kukusebenzisa I-Visual Studio 2022 okanye ngaphezulu ngomthwalo womsebenzi wophuhliso lwedesktop ye-.NET ovuliweyo. Ukusuka apho ungenza iiprojekthi Ii-apps zeWinUI 3, WPF, .NET MAUI okanye ii-console kwi-C# okanye kwi-VB.NET ebhekisa kwi-ONNX Runtime kunye neWindows ML.
Ngokuphathelele ukuxhomekeka, kukho iipakeji eziphambili zeNuGet eziphindaphindwayo kwiimeko ezininzi: I-Microsoft.ML.OnnxRuntime ye-CPU, IMicrosoft.ML.OnnxRuntime.DirectML Ukuba ufuna ukusebenzisa iGPU ngeDirectML, IMicrosoft.AI.MachineLearning ukusebenza neWindows ML kwiWinUI 3, kunye neelayibrari ezifana I-SixLabors.ImageSharp okanye okufanayo nangaphambi kokucubungula imifanekiso yokufaka.
Kakade ke, kuya kufuneka ubuncinci enye imodeli ye-ONNX ehambelanayoUngakhuphela iimodeli eziqeqeshwe kwangaphambili kwi ONNX Model Zoo (umzekelo, iResNet, iSqueezeNet, iimodeli zokwahlulahlula, ukufunyanwa kwezinto, okanye iNLP), okanye uguqule iimodeli zakho kwiPyTorch okanye kwiTensorFlow usebenzisa izixhobo ezifana Isixhobo se-AI se-VS Code, i-tf2onnx okanye i-onnxruntime-tools.
Ngokuphathelele izixhobo zekhompyutha, nangona i-ONNX Runtime iphuculiwe yi-CPU kwaye isebenza kakuhle ngaphandle kwe-GPU, uya kubona ukuphucuka okucacileyo xa uyisebenzisa. izikhawulezisi zasekuhlaleni ezifana neGPU okanye i-NPUingakumbi kwiimodeli ezinkulu (iimodeli zolwimi, i-Stable Diffusion, i-Whisper, njl.njl.). Jonga eyethu uluhlu lokuhlola oluyimfuneko lwezixhobo.
Indlela yokufaka nokusebenzisa i-ONNX Runtime kwiiprojekthi ze-.NET kunye ne-WinUI 3
Enye yezona ndlela zithe ngqo zokuqalisa Ukusebenzisa i-ONNX Runtime kwiWindows 11 Imalunga nokudala usetyenziso lwedesktop kwi-.NET. Ipateni esisiseko iyaphindwa kwi-C# nakwi-VB.NET: bhekisa kwiphakheji, ulayishe imodeli, ulungiselele amagalelo, kwaye usebenzise iseshoni yoqikelelo.
Kwiprojekthi yesiqhelo ye-.NET (umzekelo, usetyenziso lwekhonsoli okanye usetyenziso lwedesktop oluqhelekileyo), indlela eqhelekileyo kukongeza iphakheji I-Microsoft.ML.OnnxRuntime kwi-Visual Studio NuGet Package Manager, uze emva koko wenze i Iseshoni yoQhagamshelwano ukuqala kwindlela eya kwifayile yakho ye-.onnx. Loo seshoni yileyo uza kuyisebenzisa kwakhona ngalo lonke ixesha ufuna ukwenza uqikelelo.
Kwimeko yesicelo Idesktop yeWinUI 3 Nge-Windows App SDK, inkqubo iyafana, kodwa ngokuqhelekileyo uya kudibanisa i-ONNX Runtime nezinye iilayibrari ezixhasayo. Umzekelo, idemo eqhelekileyo yokwahlulahlula imifanekiso ingabandakanya:
- I-Microsoft.ML.OnnxRuntime kunye ne-Microsoft.ML.OnnxRuntime.Managed ukuphunyezwa komzekelo.
- I-SixLabors.ImageSharp ukulayisha, ukulinganisa ubungakanani kunye nokwenza imifanekiso ibe ngokwesiqhelo.
- I-SharpDX.DXGI ukuba ufuna ukukhetha ngokucacileyo iadaptha yemizobo yeDirectML.
Ngaphakathi kwifestile ephambili (umzekelo, I-MainWindow.xaml.cs) uhlala udala indlela yokuqalisa imodeli elungiselela i Iinketho zeSessionUmboneleli wokwenza (i-CPU okanye i-DirectML) kunye nendlela eya kwimodeli ye-ONNX. Ngale ndlela, usetyenziso lwakho luqala, luqalise iseshoni kube kanye kuphela, kwaye emva koko luphinde lusebenzise loo seshoni kwifowuni nganye yokubonisa, nto leyo eyenza ukuba ukusebenza kube ngcono.
Ukuba ukhetha ukusebenza kunye I-Windows ML ngqo Kwiprojekthi yeWinUI 3, uza kongeza iphakheji IMicrosoft.AI.MachineLearning kwaye uza kukopa imodeli yakho ye-ONNX (umzekelo, i-model_mnist.onnx okanye imodeli yokwahlulahlula imifanekiso) kwifolda efana Iiasethi/ML kwaye uza kumisela iipropati zayo ukuba zifakwe kwisiphumo sokuhlanganiswa njengomxholo.
Umzekelo osebenzayo: Uhlu lwemifanekiso nge-ONNX Runtime kunye ne-DirectML
Umzekelo oqhelekileyo nobonisa kakuhle I-AI yasekhaya ene-ONNX Runtime kwiWindows 11 Luhlelo lwemifanekiso kusetyenziswa imodeli efana nale I-ResNet50 okanye i-SqueezeNetUkuhamba kwengcamango kuhlala kufana, nangona ikhowudi inokwahluka ngokuxhomekeke kwisakhelo osisebenzisayo.
Kwi-app yeWinUI 3, unokuyila ujongano olulula nge iqhosha lokukhetha ifotoulawulo Umfanekiso wokubonisa umfanekiso okhethiweyo kunye I-TextBlock ukuze idwelise uqikeleloUmsebenzisi ucofa iqhosha, iFayileOpenPicker iyavula, umfanekiso ukhethwe, kwaye ngasemva, inkqubo iyawucubungula ize iwudlulisele kwimodeli ye-ONNX.
Ukulungiswa kwangaphambili komfanekiso kudla ngokubandakanya yilayishe kwifomathi ye-24-bit RGByitshintshe ubungakanani bayo ibe kubungakanani obulindelwe yimodeli (umzekelo, iiphikseli ezingama-224×224), yinqumle ukuba kuyimfuneko, kwaye uyilungelelanise ixabiso letshaneli nganye ngokweendlela kunye nokuphambuka okuqhelekileyo okuboniswe kuxwebhu lwemodeli. Nge-ImageSharp, oku kuthetha ukutshintsha ubungakanani bomfanekiso, ukuphindaphinda umqolo ngomqolo weephikseli, kunye nokugcwalisa i- I-DenseTensor ngamaxabiso aqhelekileyo.
Emva kokulungisa i-tensioner, Ixabiso le-Ort elivela kwimemori ye-tensor (ngaphandle kweekopi ezingafunekiyo), kwakhiwa isichazi-magama esinee-entry zemodeli (ngesiqhelo i-tensor enye ebizwa ngokuba yinto efana ne "data" okanye efanayo) kwaye indlela iyabizwa Run ukusuka kwi-InferenceSession. Isiphumo sidla ngokuba yi-probability tensor ngeklasi nganye, onokuyicwangcisa emva koko usebenzisa i-softmax ukuze amaxabiso abe kwi-1 kwaye afikelele kwi-1.
Okokugqibela, ucwangcisa amathuba ukusuka kwelona liphezulu ukuya kwelona liphantsi, uze ushiyeke Iilebhile ezili-10 ezinokuzithemba okungakumbi kwaye uzibonise kwiscreen. Oku kudla ngokwenziwa kusetyenziswa iklasi encedisayo (umzekelo, iipropati zePrediction with Label and Confidence) kunye netafile yeleyibhile (LabelMap) edibanisa isalathisi ngasinye semveliso kwiklasi ye-semantic (inja, imoto, ikhibhodi, njl.njl.).
Olu hlobo lomzekelo lubonisa indlela i-ONNX Runtime, i-DirectML, kunye ne-WinUI 3 ezidibana ngayo ukuze zinike Fumana amava e-AI yasekuhlaleni ene-UI yanamhlanje, ukukhawuleziswa kwehardware, kwaye akukho mfuneko yoqhagamshelo lwe-intanethi oluqhubekayo..
Ukusebenza ngeemodeli zeWindows ML kunye ne-ONNX kwiWinUI 3
Ukongeza ekusebenziseni i-"pure" ONNX Runtime, ungathembela kwi I-Windows ML ngaphakathi kweprojekthi yeWinUI 3 ukuze uzuze ekuhlanganisweni kwenkqubo yokusebenza yendalo. Le ndlela iluncedo kakhulu xa ufuna ukupakisha usetyenziso lwakho kwi-MSIX, usebenzise iWindows App SDK, kwaye uqinisekise ukuba iyasebenza rhoqo kwizixhobo zanamhlanje zeWindows 11.
Inkqubo eqhelekileyo ibandakanya Yongeza iphakheji yeMicrosoft.AI.MachineLearning kwiprojekthi, kopisha imodeli yakho ye-ONNX (umzekelo, i-model_mnist.onnx okanye imodeli yokwahlulahlula imifanekiso) kwifolda efana Iiasethi/ML kwaye ulungiselele iipropati zayo ukuba zifakwe kwisiphumo sokuhlanganiswa njengomxholo.
Ukusuka kwikhowudi ye-C#, ulayisha imodeli usebenzisa ii-Windows ML APIs, Udala iseshoni kwaye ulungiselela idatha yokufaka kwifomathi elindelekileyo (umzekelo, i-tensor enomfanekiso we-28×28 we-grayscale we-MNIST, okanye umfanekiso we-224×224 wemodeli yombono ontsonkothileyo). Emva koko iseshoni isebenza ngokusebenzisa Ukukhawuleziswa kwe-DirectML kunye nehardware ngokucacileyo ukuba isixhobo siyayixhasa.
Nje ukuba uqikelelo lwenziwe, usebenza ngeemveliso ukuze ufumane ixabiso elinokwenzeka kakhulu (umzekelo, inani elinokwenzeka kakhulu kwimeko ye-MNIST) okanye iseti yamathuba onokuwabonisa kwi-interface. Konke oku kudibana ngokungenamthungo ne I-WinUI 3 XAML UI, ekuvumela ukuba uhlaziye ulawulo olubonakalayo ngeziphumo ze-AI.
I-Windows ML ikwaphatha ababoneleli bokusebenza kunye lungelelanisa ukusetyenziswa kwe-CPU, i-GPU, kunye ne-NPU kuxhomekeke kwizixhobo ezikhoyo, ngaphandle kokutshintsha iimodeli zakho okanye ukulungelelanisa ngokutsha ikhowudi yakho.
Iimodeli zolwimi zasekuhlaleni, iiRAG, kunye neminye imizekelo ephucukileyo
I-AI yengingqi kwiWindows 11 ayipheleli nje ekuboniseni ikhompyutha. Nge-ONNX Runtime kunye ne-DirectML, ungayisebenzisa Iimodeli zolwimi (LLM) ziguqulelwe kwi-ONNX, iimodeli zelizwi ukuya kwisicatshulwa ezifana neWhisper, okanye ukwahlulahlula kunye nokwakha imifanekiso efana neStable Diffusion.
IMicrosoft igcina inani le imizekelo esemthethweni kunye ne-AI Dev GalleryUsetyenziso oluvulekileyo oluneedemo ezingaphezu kwama-25 ezisebenzisanayo ezibonisa indlela yokudibanisa iimodeli zasekuhlaleni kwii-apps zokwenyani: ukusuka kubahleli besandi abasebenzisa i-AI abadwelisa iziqwenga ngokusekelwe kumbuzo wombhalo, ukuya kwii-apps ezikrelekrele zokuthatha amanqaku ezisebenzisa iimodeli ezifana Phi-3 ukuze kufinyezwe isishwankathelo, kugqitywe ngokuzenzekelayo, kwaye kuqiqwe ngaphandle kwe-intanethi.
Ipateni efanelekileyo kakhulu yile IsiZukulwana esomeleziwe soBuyiselo (RAG)Kule ndlela, udibanisa imodeli yolwimi lwasekuhlaleni kunye nolwazi lwangaphandle. Umzekelo, i-app ye-WPF ingalayisha i-PDF, iyinqumle, kwaye iyifake kwi-vector isebenzisa imodeli yokufakela yasekuhlaleni, igcine ezo vector, kwaye, ngexesha lombuzo, ifumane iziqwenga ezifanelekileyo ukuze imodeli yolwimi ikwazi ukuvelisa iimpendulo ezisekelwe kwidatha yokwenyani, ngaphandle kokutshintsha ubunzima bemodeli.
Le mizekelo ibonisa ukuba ungakha abancedisi abasebenzayo beemeko ezithile, iinkqubo zokukhangela intsingiselo, ukubhalwa kwesandi, kunye nee-chatbots ezikhethekileyo. yasekhaya ngokupheleleyo kwiWindows 11, kusetyenziswa i-ONNX Runtime kunye ne-DirectML ukugcina ukusebenza kuphantsi kolawulo.
Ukongeza, kukho iiprojekthi zesalathiso zokusebenzisa iimodeli ezinje I-Phi-3, iFlame 3 okanye iMistral Ilungiselelwe ngeDirectML, zombini kwifomathi ye-ONNX kwaye ngqo kwiPyTorch, kusetyenziswa usetyenziso olukhaphukhaphu (umz., nge-interface yewebhu yesitayile seGradio) ukuvavanya ukusebenza kunye nokulungisa useto ngaphambi kokuba ludityaniswe kwisicelo sokugqibela.
I-AI ekhawulezisiweyo kwi-Hardware kwi-intanethi nge-ONNX Runtime Web kunye ne-WebNN
Enye indlela enomdla yokusebenzisa I-AI yasekhaya ene-ONNX Runtime kwiWindows 11 Yindawo yewebhu, ekuxhasayo I-ONNX Web Runtime kunye ne-WebNN APIKule meko, iimodeli ezinzima ezifana neStable Diffusion, Segment Anything, okanye Whisper zinokuqhutywa ngqo kwiGPU okanye kwi-NPU yesixhobo ngeDirectML, kodwa zivela kwi-web application yanamhlanje.
Kukho imiboniso kawonke-wonke ebonisa, umzekelo, I-Stable Diffusion Turbo evelisa imifanekiso evela kwisicatshulwa kwimizuzwana embalwa nje usebenzisa iWebNN kunye neDirectML kwiimashini ezinezakhono ze-AI eziphambili. Oku kuyasebenza nakwi- Icandelo Nantoni naekuvumela ukuba unciphise nayiphi na into kumfanekiso olayishwe ngumsebenzisi, okanye nge Isiseko seWhisper, eguqula intetho ibe ngumbhalo kwindawo ekuyo kwisiphequluli.
Kwezi meko, i-ONNX Runtime Web iphatha layisha imodeli ye-ONNX kumxholo wesikhangelilawula ii-tensors zokufaka kunye neziphumo kwaye unxibelelane neWebNN ukuze uhambise ukusebenza kwiDirectML kwisixhobo seWindows 11. Isiphumo kukuba ungazisa amava e-AI atyebileyo ngaphandle kokuthumela isandi, imifanekiso, okanye umbhalo kwiseva ekude.
Olu dibaniso lwenza iWindows 11 ibe liqonga elinamandla kakhulu kungekuphela nje usetyenziso lwedesktop lwendalo, kodwa kunye ne usetyenziso lwewebhu olunengqiqo exutyiweyoapho inxalenye yengqiqo ihlala kwilifu kwaye inxalenye isetyenziswa ngqo kwikhompyutha yomsebenzisi.
Kwabaphuhlisi abasebenzisa iteknoloji zewebhu (iJavaScript, ii-SPA frameworks, njl.njl.), le ndlela yindlela elula yokwenza ngaphandle kwesidingo sokubhala ngokutsha lonke usetyenziso Yiya kwi-AI yasekuhlaleni ngaphandle kokuphinda ubhale yonke i-application kwi-.NET okanye kwi-C#.
Iindlela ezilungileyo, ukhuseleko kunye nophuhliso olunenkathalo
Ngokudibanisa Iimpawu ze-AI kwizicelo zeWindowsIMicrosoft icebisa ukuba kulandelwe uthotho lwezikhokelo zophuhliso ezinoxanduva, ingakumbi kwimeko yeemodeli ezivelisayo kunye nokucubungula umxholo obuthathaka. Oku akupheleli nje ekusebenzeni, kodwa kukwamalunga nokunciphisa ingozi okanye ukusetyenziswa gwenxa.
Ii-API zeWindows AI ziquka iindlela zokusebenza ukumodareyitha umxholo wombhalo ngokusebenzisa iinkonzo ezifana neMicrosoft Foundry, enceda ekuhluzeni iziphumo ezinokuba yingxaki. Nokuba xa usebenza neemodeli zasekuhlaleni, kuyacetyiswa ukusebenzisa ezinye iindlela zokuqinisekisa okokufaka kunye nokuhlolwa kwemveliso kwiindawo ezibalulekileyo (ezemali, ezempilo, ukhuseleko, njl.njl.).
Ngokwembono yoyilo lwesicelo, kuyacetyiswa qinisekisa iziphumo ngaphambi kokuba uqalise uqikelelo (ubungakanani bomfanekiso, iintlobo zeefayile, ubude bombhalo), lawula ukuphahlazeka kwe-GPU okanye ukungabikho kokukhawulezisa ngeendlela ezahlukeneyo ze-CPU, kwaye uphephe ukulayisha imodeli ngokuphindaphindiweyo kwisicelo ngasinye. Ukugcina iseshoni yoqikelelo iphila kwaye uyisebenzise kwakhona kubalulekile ekusebenzeni kakuhle.
Kwakhona kubalulekile ukuqwalasela ukwenziwa ngcono kweemodeli ze-ONNXIindlela ezinje ngokulinganisa ubungakanani, ukuthena, kunye nokudibanisa ukusebenza kunokunciphisa ubungakanani bemodeli kwaye kuphucule isantya ngaphandle kokunciphisa ukuchaneka okukhulu. Izixhobo ezifana ne-AI Toolkit ye-VS Code okanye isiseko sokwakha se-ONNX Runtime (kuquka ii-API ezintsha zokwakha ezivezwe kwiinguqulelo zamva nje) zenzelwe ukwenza lula le nkqubo.
Okokugqibela, ungalibali ukujonga ukufumaneka kwehadiwe kunye nemo yomphuhlisi kwisixhobo apho uza kusebenzisa khona ii-apps zakho, ingakumbi ukuba uza kuvavanya izixhobo zabasebenzisi bokugqibela, ii-kiosk, izixhobo zoshishino, okanye iindawo ezinemigaqo-nkqubo yokhuseleko engqongqo.
Njengoko zonke ezi zinto zilungelelaniswe kakuhle—i-ONNX njengefomathi eqhelekileyo, i-ONNX Runtime kunye neWindows ML njengeenjini zokuphumeza, i-DirectML kunye nee-EP ezivela kwi-AMD, i-Intel, i-NVIDIA, kunye ne-Qualcomm ukuze zisebenzise i-CPU, i-GPU, kunye ne-NPU, kunye nezixhobo ezifana ne-AI Dev Gallery okanye i-AI Toolkit—iWindows 11 iba liqonga elinamandla kakhulu lokwakha. Izicelo ze-AI zasekhaya, zabucala, ezikhawulezayo, nezilungele imveliso, ukusuka kwiindlela ezilula zokwahlulahlula imifanekiso ukuya kwiindlela ezintsonkothileyo ezisekelwe kwiimodeli zolwimi kunye neRAG.
