【GDE 分享】移植 Mediapipe LLM Demo 到 Kotlin Multiplatform

// app/src/commonMain/.../llm/LLMOperatorinterface LLMOperator {
    /**     * To load the model into current context.     * @return 1. null if it went well 2. an error message in string     */    suspend fun initModel(): String?
    fun sizeInTokens(text: String): Int
    suspend fun generateResponse(inputText: String): String
    suspend fun generateResponseAsync(inputText: String): Flow<Pair<String, Boolean>>
}

class LLMInferenceAndroidImpl(private val ctx: Context): LLMOperator {
    private lateinit var llmInference: LlmInference    private val initialized = AtomicBoolean(false)    private val partialResultsFlow = MutableSharedFlow<Pair<String, Boolean>>(...)
    override suspend fun initModel(): String? {        if (initialized.get()) {            return null        }        return try {            val modelPath = ...            if (File(modelPath).exists().not()) {                return "Model not found at path: $modelPath"            }            loadModel(modelPath)            initialized.set(true)            null        } catch (e: Exception) {            e.message        }    }    private fun loadModel(modelPath: String) {        val options = LlmInference.LlmInferenceOptions.builder()            .setModelPath(modelPath)            .setMaxTokens(1024)            .setResultListener { partialResult, done ->                // Transforming the listener to flow,                // making it easy on UI integration.                partialResultsFlow.tryEmit(partialResult to done)            }            .build()
        llmInference = LlmInference.createFromOptions(ctx, options)    }
    override fun sizeInTokens(text: String): Int = llmInference.sizeInTokens(text)
    override suspend fun generateResponse(inputText: String): String {        ...        return llmInference.generateResponse(inputText)    }
    override suspend fun generateResponseAsync(inputText: String): Flow<Pair<String, Boolean>> {        ...        llmInference.generateResponseAsync(inputText)        return partialResultsFlow.asSharedFlow()    }
}

// app/build.gradle.ktsplugins {    ...    alias(libs.plugins.cocoapods)}cocoapods {    ...    ios.deploymentTarget = "15"
    pod("MediaPipeTasksGenAIC") {        version = "0.10.14"        extraOpts += listOf("-compiler-option", "-fmodules")    }    pod("MediaPipeTasksGenAI") {        version = "0.10.14"        extraOpts += listOf("-compiler-option", "-fmodules")    }}

// 注意这些 import 是 cocoapods 开头的import cocoapods.MediaPipeTasksGenAI.MPPLLMInferenceimport cocoapods.MediaPipeTasksGenAI.MPPLLMInferenceOptionsimport platform.Foundation.NSBundle...class LLMOperatorIOSImpl: LLMOperator {
    private val inference: MPPLLMInference
        init {        val modelPath = NSBundle.mainBundle.pathForResource(..., "bin")
        val options = MPPLLMInferenceOptions(modelPath!!)        options.setModelPath(modelPath!!)        options.setMaxTokens(2048)        options.setTopk(40)        options.setTemperature(0.8f)        options.setRandomSeed(102)
        // NPE was thrown here right after it printed the success initialization message internally.        inference = MPPLLMInference(options, null)     }
    override fun generateResponse(inputText: String): String {...}    override fun generateResponseAsync(inputText: String, ...) :... {        ...    }    ...}

constructor(  options: cocoapods.MediaPipeTasksGenAI.MPPLLMInferenceOptions,   error: CPointer<ObjCObjectVar<platform.Foundation.NSError?>>?)

// 其中一种尝试memScoped {    val pp: CPointerVar<ObjCObjectVar<NSError?>> = allocPointerTo()    val inference = MPPLLMInference(options, pp.value)    Napier.i(pp.value.toString())}

// 1. 声明一个类似 LLMOperator 的接口但更简单，方便适配 iOS 的 SDK。// app/src/iosMain/.../llm/LLMOperator.ktinterface LLMOperatorSwift {    suspend fun loadModel(modelName: String)    fun sizeInTokens(text: String): Int    suspend fun generateResponse(inputText: String): String    suspend fun generateResponseAsync(        inputText: String,        progress: (partialResponse: String) -> Unit,        completion: (completeResponse: String) -> Unit    )}
// 2. 在 iOS 工程里实现这个接口// iosApp/iosApp/LLMInferenceDelegate.swiftclass LLMOperatorSwiftImpl: LLMOperatorSwift {    ...    var llmInference: LlmInference?
    func loadModel(modelName: String) async throws {        let path = Bundle.main.path(forResource: modelName, ofType: "bin")!        let llmOptions =  LlmInference.Options(modelPath: path)        llmOptions.maxTokens = 4096        llmOptions.temperature = 0.9
        llmInference = try LlmInference(options: llmOptions)    }
    func generateResponse(inputText: String) async throws -> String {        return try llmInference!.generateResponse(inputText: inputText)    }
    func generateResponseAsync(inputText: String, progress: @escaping (String) -> Void, completion: @escaping (String) -> Void) async throws {        try llmInference!.generateResponseAsync(inputText: inputText) { partialResponse, error in            // progress            if let e = error {                print("\(self.errorTag) \(e)")                completion(e.localizedDescription)                return            }            if let partial = partialResponse {                progress(partial)            }        } completion: {            completion("")        }    }    ...    }
// 3. iOS 再把代理好的（重点是初始化）类传回给 Kotlin// iosApp/iosApp/iosApp.swiftclass AppDelegate: UIResponder, UIApplicationDelegate {    ...    func application(）{        ...        let delegate = try LLMOperatorSwiftImpl()        MainKt.onStartup(llmInferenceDelegate: delegate)            }}
// 4. 最初 iOS 在 KMP 上的实现细节直接代理给该对象（通过构造器注入）class LLMOperatorIOSImpl(   private val delegate: LLMOperatorSwift) : LLMOperator {      ...}

// Commonexpect class LLMOperatorFactory {    fun create(): LLMOperator}val sharedModule = module {   // 从不同的 LLMOperatorFactory 创建出 Common 层所需的 LLMOperator  single<LLMOperator> { get<LLMOperatorFactory>().create() }}
// Androidactual class LLMOperatorFactory(private val context: Context){    actual fun create(): LLMOperator = LLMInferenceAndroidImpl(context)}val androidModule = module {    // Android 注入 App 的 Context    single { LLMOperatorFactory(androidContext()) }}
// iOSactual class LLMOperatorFactory(private val llmInferenceDelegate: LLMOperatorSwift) {    actual fun create(): LLMOperator = LLMOperatorIOSImpl(llmInferenceDelegate)}
module {    // iOS 注入 onStartup 函数传入的 delegate    single { LLMOperatorFactory(llmInferenceDelegate) }}

// Android 版本class ChatViewModel(    private val inferenceModel: InferenceModel) : ViewModel() {...}
// KMP 版本，转换 ViewModel 为 ScreenModel，并修改传入对象class ChatViewModel(    private val llmOperator: LLMOperator) : ScreenModel {...}

// Android 版本@Composableinternal fun ChatRoute(    chatViewModel: ChatViewModel = viewModel(        factory = ChatViewModel.getFactory(LocalContext.current.applicationContext)    )) {    ...    ChatScreen(...) {...}}
// KMP 版本，改成外部初始化后传入@Composableinternal fun ChatRoute(    chatViewModel: ChatViewModel) {
// 此处采用了默认参数注入的方案，便于解耦。// koinInject() 是 Koin 官方提供的针对 Compose // 的 @Composable 函数注入的一个方法。@Composablefun AiScreen(llmOperator:LLMOperator = koinInject()) {    // 使用 ScreenModel 的 remember 方法    val chatViewModel = rememberScreenModel { ChatViewModel(llmOperator) }    ...    Column {        ...        Box(...) {            if (showLoading) {                ...            } else {                ChatRoute(chatViewModel)            }        }    }}

// Android 版本inferenceModel.generateResponseAsync(fullPrompt)inferenceModel.partialResults    .collectIndexed { index, (partialResult, done) ->        ...    }
// KMP 版本，把 Flow 的返回前置了，兼容了两个平台的 SDK 设计llmOperator.generateResponseAsync(fullPrompt)    .collectIndexed { index, (partialResult, done) ->        ...    }

// Android 版本Text(stringResource(R.string.chat_label))
// KMP 版本，该引用是使用插件从 xml 映射而来// (commonMain/composeResources/values/strings.xml)import mediapiper.app.generated.resources.chat_label...Text(stringResource(Res.string.chat_label))