On-deviceModel在KMP的集成與用例

Guangzhou《Android構(gòu)建與架構(gòu)實(shí)戰(zhàn)》《面向Kotlin的Gradle基礎(chǔ)指南》2bab.me/zh/book/TableofContent

1.On-deviceModel2.KotlinMultiplatform3.Demonstration4.StrengthsofOn-deviceModelandKMPSummaryLowlatencyresponsesandenhanceddataprivacy.On-deviceModelWebEdgeComputingRight-click

andselect

‘Replaceimage’Google’smostefficientmodelforon-devicetasks.Itrunsdirectlyonmobile.GeminiNanoEdgeComputingYoucanfine-tuneGeminiNanowithLoRA(Low-RankAdaptation)toperformspecializedtasksaswellasitslargercounterparts.GeminiNanorunsinAndroid'sAICoresystemservice.GeminiNanoiscurrentlyonlyavailableonGooglePixel8Pro,Pixel8,Pixel8a,andSamsungS24Seriesdevices,withsupportformoredevicescomingsoon.GeminiNanoEdgeComputingEdgeComputingOnlyfewpartnersthatwerecooperatedwithGoogleinthisearlystage.GeminiNanoMediaPipeSolutionsprovidesasuiteoflibrariesandtoolsforyoutoquicklyapplyartificialintelligence(AI)andmachinelearning(ML)techniquesinyourapplications./homeEdgeComputingMediaPipeEdgeComputingMediaPipewithGemmaTheLLMInferenceAPIletsyourunlargelanguagemodels(LLMs)completelyon-device.ItsupportsGemma

2Band7B.Gemmaisalightweight,openmodelsbuiltfromthesame

researchandtechnologyusedtocreatetheGeminimodels.Italsosupportsthefollowingexternalmodels:Phi-2,Falcon-RW-1BandStableLM-3B.EdgeComputingGemma(1.1)2B

vs

GeminiNano

(1.8B,3.5B)Gemma2B:開源，通用性強(qiáng)，并為“使用CPU執(zhí)行”而優(yōu)化，目前尚未支持在終端設(shè)備上直接微調(diào)，通常是優(yōu)化和微調(diào)模型后，下載一個(gè)完整的新模型到終端/App。GeminiNano：非開源，需要特定芯片加速（例如Pixel8Pro的TensorG3），有更強(qiáng)大的表現(xiàn)，目前已有終端設(shè)備上直接微調(diào)的LoRA支持。都是On-deviceModel，二者技術(shù)同源。Sharingyourcodeupto99%.KotlinMultiplatform(KMP)WebMobileItwasintroducedin2017,enablesdeveloperstowritecross-platformcodeusingKotlin,sharingbusiness

logicandUI(optional)acrossAndroid(withKotlintoJVM)iOS(withKotlintoNative)Web(withKotlintoJSandWASM),DesktopincludingWindows/Linux/macOS(withKotlintoNativeandJVM)MobileKotlinMultiplatformBaseonJetpackComposeMobileFirst-classtoolingandlibrarysupportforKotlinMultiplatformonAndroid.NowexpandedtoyourfavouritelibrarieslikeRoom/DataStore.LeveragingKMPforbusinesslogicacrossAndroid,iOSandweb.StronglyenhancingtheecosystemofKMP.MobileKotlinMultiplatform

onAndroidiOSWebDesktopCaseStudiesMobileMobileKotlinMultiplatforminGoogleWorkspace

-@JasonParachoniak

KotlinConf2024MobileHitchhiker’sGuidetoKotlinMultiplatformLibraries

-@JohnO’Reilly

KotlinConf2024DemonstrationWebHands-onShowingupacommon2columnslistwithimageandtextitems.Thecodebaseforkedfrombelow:

/linroid/Pexels

@Lindroid(Itservesasatemplateforthenextdemo)Hands-onPexelsHands-onGemma2B1.1Showingupaconversationalchatinterfacewithgemma-2b-1.1models.TheUIdesignandoriginalcodebasecomefromGoogle.

/google-ai-edge/mediapipe-samples/tree/main/examples/llm_inference/androidLet’sbuildsthfun!/2BAB/Pexels/Hands-onCupertinoChoosetheMaterialdesignforanativeAndroidexperienceandtheCupertinodesignforanativeiOSexperience./alexzhirkevich/compose-cupertino/SummaryofUsecasesHands-onGenerateemails/stories/poems/jokes/lyrics/etc.Paraphrasing/GrammarRecifyingSmarttemplatesinvariousdatareports.…FocusonWriteFocusonRead/Analysis/InferenceGeneralQ&AwithknowledgebaseClassificationSummarySearchwithRAGmechanismWorkwithMultimodaltoexpressmulti-mediaresponse6.WorkwithFunctionCallingtointegratemoresystemfeatures7.WorkwithResponseFormattertogeneratestructuralresponse8.AnalyzesensitivedataHands-on未來(lái)，用戶的多數(shù)簡(jiǎn)單需求可以在本地被滿足Let’stalkaboutWHY?

StrengthsofOn-deviceModelandKMP

SummaryWhyon-devicemodel?SummaryLowCostLowLatency(Yetseenintoday’sdemo)PrivacyFluentInterop流暢的互操作

以iOS為例：C-Interop讓Kotlin獲得直接調(diào)用Swift/OC/iOS依賴的能力；反之Kotlin基于LLVM編譯而成的.framework也可以讓Swift無(wú)縫調(diào)用Kotlin寫的功能。Nati