@capgo/capacitor-llm 8.1.0 → 8.1.1

package/README.md

@@ -380,14 +380,15 @@ Get the native Capacitor plugin version.
 Model configuration options
 Only `path` is required. All other properties are optional overrides.
 
-| Prop              | Type                | Description                                                                                                                                                                                                                                                                                                                             |
-| ----------------- | ------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| **`path`**        | <code>string</code> | Model path, "Apple Intelligence" for the Apple system model on iOS, or "Gemini Nano" for the Android AICore system model when already available on the device. On iOS, custom `.litertlm` models require the plugin to be integrated through Swift Package Manager. Gemma 4 examples use `.litertlm` on mobile and `*-web.task` on web. |
-| **`modelType`**   | <code>string</code> | Optional. Model file type/extension (for example `task`, `bin`, `litertlm`, or `gemini-nano`). If not provided, it is extracted from the path. Use `litertlm` on iOS only in SwiftPM integrations.                                                                                                                                      |
-| **`maxTokens`**   | <code>number</code> | Maximum number of tokens the model handles                                                                                                                                                                                                                                                                                              |
-| **`topk`**        | <code>number</code> | Number of tokens the model considers at each step                                                                                                                                                                                                                                                                                       |
-| **`temperature`** | <code>number</code> | Amount of randomness in generation (0.0-1.0)                                                                                                                                                                                                                                                                                            |
-| **`randomSeed`**  | <code>number</code> | Optional. Random seed for generation.                                                                                                                                                                                                                                                                                                   |
+| Prop              | Type                        | Description                                                                                                                                                                                                                                                                                                                             | Since |
+| ----------------- | --------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----- |
+| **`path`**        | <code>string</code>         | Model path, "Apple Intelligence" for the Apple system model on iOS, or "Gemini Nano" for the Android AICore system model when already available on the device. On iOS, custom `.litertlm` models require the plugin to be integrated through Swift Package Manager. Gemma 4 examples use `.litertlm` on mobile and `*-web.task` on web. |       |
+| **`modelType`**   | <code>string</code>         | Optional. Model file type/extension (for example `task`, `bin`, `litertlm`, or `gemini-nano`). If not provided, it is extracted from the path. Use `litertlm` on iOS only in SwiftPM integrations.                                                                                                                                      |       |
+| **`maxTokens`**   | <code>number</code>         | Maximum number of tokens the model handles                                                                                                                                                                                                                                                                                              |       |
+| **`topk`**        | <code>number</code>         | Number of tokens the model considers at each step                                                                                                                                                                                                                                                                                       |       |
+| **`temperature`** | <code>number</code>         | Amount of randomness in generation (0.0-1.0)                                                                                                                                                                                                                                                                                            |       |
+| **`randomSeed`**  | <code>number</code>         | Optional. Random seed for generation.                                                                                                                                                                                                                                                                                                   |       |
+| **`backend`**     | <code>'gpu' \| 'cpu'</code> | Optional. LiteRT-LM engine backend for iOS (SwiftPM) and Android. Use `cpu` for stable long generations. When omitted, iOS prefers CPU then falls back to GPU; Android uses CPU.                                                                                                                                                        | 8.2.0 |
 
 
 #### DownloadModelResult

package/android/src/main/java/ee/forgr/capgo_llm/LLM.java

@@ -84,9 +84,10 @@ public class LLM {
         Integer topk,
         Float temperature,
         Integer randomSeed,
+        String backend,
         ModelLoadCallback callback
     ) {
-        ModelLoadRequest request = nextLoadRequest(path, modelType, maxTokens, topk, temperature, randomSeed);
+        ModelLoadRequest request = nextLoadRequest(path, modelType, maxTokens, topk, temperature, randomSeed, backend);
         resetModelState();
 
         android.util.Log.d(
@@ -111,10 +112,11 @@ public class LLM {
         Integer maxTokens,
         Integer topk,
         Float temperature,
-        Integer randomSeed
+        Integer randomSeed,
+        String backend
     ) {
         currentLoadId += 1;
-        return new ModelLoadRequest(path, modelType, maxTokens, topk, temperature, randomSeed, currentLoadId);
+        return new ModelLoadRequest(path, modelType, maxTokens, topk, temperature, randomSeed, backend, currentLoadId);
     }
 
     private synchronized boolean isCurrentLoad(long loadId) {
@@ -210,11 +212,12 @@ public class LLM {
     }
 
     private Engine initializeLiteRtModel(String actualPath, ModelLoadRequest request) {
+        Backend primaryBackend = resolveLiteRtBackend(request.backend);
         EngineConfig config = new EngineConfig(
             actualPath,
-            new Backend.CPU(),
-            new Backend.CPU(),
-            new Backend.CPU(),
+            primaryBackend,
+            primaryBackend,
+            primaryBackend,
             request.maxTokens,
             context.getCacheDir().getAbsolutePath()
         );
@@ -223,6 +226,14 @@ public class LLM {
         return loadedEngine;
     }
 
+    private Backend resolveLiteRtBackend(String backend) {
+        if (backend != null && "gpu".equalsIgnoreCase(backend.trim())) {
+            return new Backend.GPU();
+        }
+
+        return new Backend.CPU();
+    }
+
     private LlmInference initializeMediaPipeModel(String actualPath, ModelLoadRequest request) {
         LlmInferenceOptions options = LlmInferenceOptions.builder()
             .setModelPath(actualPath)
@@ -820,6 +831,7 @@ public class LLM {
         private final Integer topk;
         private final Float temperature;
         private final Integer randomSeed;
+        private final String backend;
         private final long loadId;
 
         private ModelLoadRequest(
@@ -829,6 +841,7 @@ public class LLM {
             Integer topk,
             Float temperature,
             Integer randomSeed,
+            String backend,
             long loadId
         ) {
             this.modelPath = modelPath;
@@ -837,6 +850,7 @@ public class LLM {
             this.topk = topk;
             this.temperature = temperature;
             this.randomSeed = randomSeed;
+            this.backend = backend;
             this.loadId = loadId;
         }
     }

package/android/src/main/java/ee/forgr/capgo_llm/LLMPlugin.java

@@ -9,7 +9,7 @@ import com.getcapacitor.annotation.CapacitorPlugin;
 @CapacitorPlugin(name = "CapgoLLM")
 public class LLMPlugin extends Plugin {
 
-    private final String pluginVersion = "8.1.0";
+    private final String pluginVersion = "8.1.1";
 
     private LLM llm;
 
@@ -106,6 +106,7 @@ public class LLMPlugin extends Plugin {
         Float temperature = call.getFloat("temperature", 0.8f);
         Integer randomSeed = call.getInt("randomSeed", 0);
         String modelType = call.getString("modelType");
+        String backend = call.getString("backend");
 
         llm.setModel(
             path,
@@ -114,6 +115,7 @@ public class LLMPlugin extends Plugin {
             topk,
             temperature,
             randomSeed,
+            backend,
             new LLM.ModelLoadCallback() {
                 @Override
                 public void onSuccess() {

package/dist/docs.json

@@ -374,6 +374,18 @@
           "docs": "Optional. Random seed for generation.",
           "complexTypes": [],
           "type": "number | undefined"
+        },
+        {
+          "name": "backend",
+          "tags": [
+            {
+              "text": "8.2.0",
+              "name": "since"
+            }
+          ],
+          "docs": "Optional. LiteRT-LM engine backend for iOS (SwiftPM) and Android.\nUse `cpu` for stable long generations. When omitted, iOS prefers CPU then falls back to GPU; Android uses CPU.",
+          "complexTypes": [],
+          "type": "'gpu' | 'cpu' | undefined"
         }
       ]
     },

package/dist/esm/definitions.d.ts

@@ -188,4 +188,10 @@ export interface ModelOptions {
     temperature?: number;
     /** Optional. Random seed for generation. */
     randomSeed?: number;
+    /**
+     * Optional. LiteRT-LM engine backend for iOS (SwiftPM) and Android.
+     * Use `cpu` for stable long generations. When omitted, iOS prefers CPU then falls back to GPU; Android uses CPU.
+     * @since 8.2.0
+     */
+    backend?: 'gpu' | 'cpu';
 }

package/dist/esm/definitions.js.map

@@ -1 +1 @@
-{"version":3,"file":"definitions.js","sourceRoot":"","sources":["../../src/definitions.ts"],"names":[],"mappings":"","sourcesContent":["/**\n * LLM Plugin interface for interacting with on-device language models\n */\nexport interface LLMPlugin {\n  /**\n   * Creates a new chat session\n   * @returns Promise with chat id and optional instructions\n   */\n  createChat(): Promise<{ id: string; instructions?: string }>;\n\n  /**\n   * Sends a message to the AI in a specific chat session\n   * @param options - The chat id and message to send\n   * @returns Promise that resolves when message is sent\n   */\n  sendMessage(options: { chatId: string; message: string }): Promise<void>;\n\n  /**\n   * Gets the readiness status of the LLM\n   * @returns Promise with readiness status string\n   */\n  getReadiness(): Promise<{ readiness: string }>;\n\n  /**\n   * Sets the model configuration\n   * - iOS: Use \"Apple Intelligence\" as path for the system model. Custom LiteRT-LM `.litertlm` models are supported on iOS only when this plugin is integrated through Swift Package Manager.\n   * - Android: Use \"Gemini Nano\" for the AICore system model on supported devices where it is already available. Prefer LiteRT-LM `.litertlm` bundles for custom models; legacy MediaPipe `.task` models are still supported\n   * - Web: Provide a web-ready model asset for `@mediapipe/tasks-genai` such as Gemma 4 `*-web.task`\n   * @param options - The model configuration\n   * @returns Promise that resolves when model is loaded\n   */\n  setModel(options: ModelOptions): Promise<void>;\n\n  /**\n   * Downloads a model from a URL and saves it to the appropriate location\n   * - iOS: Downloads to the app's documents directory\n   * - Android: Downloads to the app's files directory\n   * @param options - The download configuration\n   * @returns Promise with download result including the saved path\n   */\n  downloadModel(options: DownloadModelOptions): Promise<DownloadModelResult>;\n\n  /**\n   * Adds a listener for text received from AI\n   * @param eventName - Event name 'textFromAi'\n   * @param listenerFunc - Callback function for text events\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'textFromAi',\n    listenerFunc: (event: TextFromAiEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Adds a listener for AI completion events\n   * @param eventName - Event name 'aiFinished'\n   * @param listenerFunc - Callback function for finish events\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'aiFinished',\n    listenerFunc: (event: AiFinishedEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Adds a listener for generation failures that happen after streaming starts\n   * @param eventName - Event name 'generationError'\n   * @param listenerFunc - Callback function for generation errors\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'generationError',\n    listenerFunc: (event: GenerationErrorEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Adds a listener for model download progress events\n   * @param eventName - Event name 'downloadProgress'\n   * @param listenerFunc - Callback function for progress events\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'downloadProgress',\n    listenerFunc: (event: DownloadProgressEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Adds a listener for readiness status changes\n   * @param eventName - Event name 'readinessChange'\n   * @param listenerFunc - Callback function for readiness events\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'readinessChange',\n    listenerFunc: (event: ReadinessChangeEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Get the native Capacitor plugin version.\n   *\n   * @returns Promise that resolves with the plugin version\n   * @throws Error if getting the version fails\n   * @since 1.0.0\n   * @example\n   * ```typescript\n   * const { version } = await CapgoLLM.getPluginVersion();\n   * console.log('Plugin version:', version);\n   * ```\n   */\n  getPluginVersion(): Promise<{ version: string }>;\n}\n\n/**\n * Event data for text received from AI\n */\nexport interface TextFromAiEvent {\n  /** The text content from AI - this is an incremental chunk, not the full text */\n  text: string;\n  /** The chat session ID */\n  chatId: string;\n  /** Whether this is a complete chunk (true) or partial streaming data (false) */\n  isChunk?: boolean;\n}\n\n/**\n * Event data for AI completion\n */\nexport interface AiFinishedEvent {\n  /** The chat session ID that finished */\n  chatId: string;\n}\n\n/**\n * Event data for generation failures\n * `chatId` is optional and may be omitted when the failure is not tied to a specific chat.\n */\nexport interface GenerationErrorEvent {\n  /** Optional. The chat session ID that failed, when available. */\n  chatId?: string;\n  /** Error message describing the failure */\n  error: string;\n}\n\n/**\n * Options for downloading a model\n * Only `url` is required. `companionUrl` and `filename` are optional.\n */\nexport interface DownloadModelOptions {\n  /** URL of the model file to download */\n  url: string;\n  /** Optional: URL of a companion file for legacy model formats */\n  companionUrl?: string;\n  /** Optional: Custom filename (defaults to filename from URL) */\n  filename?: string;\n}\n\n/**\n * Result of model download\n */\nexport interface DownloadModelResult {\n  /** Path where the model was saved */\n  path: string;\n  /** Path where the companion file was saved (if applicable) */\n  companionPath?: string;\n}\n\n/**\n * Event data for download progress\n */\nexport interface DownloadProgressEvent {\n  /** Percentage of download completed (0-100) */\n  progress: number;\n  /** Total bytes to download */\n  totalBytes?: number;\n  /** Bytes downloaded so far */\n  downloadedBytes?: number;\n}\n\n/**\n * Event data for readiness status changes\n */\nexport interface ReadinessChangeEvent {\n  /** The readiness status */\n  readiness: string;\n}\n\n/**\n * Model configuration options\n * Only `path` is required. All other properties are optional overrides.\n */\nexport interface ModelOptions {\n  /** Model path, \"Apple Intelligence\" for the Apple system model on iOS, or \"Gemini Nano\" for the Android AICore system model when already available on the device. On iOS, custom `.litertlm` models require the plugin to be integrated through Swift Package Manager. Gemma 4 examples use `.litertlm` on mobile and `*-web.task` on web. */\n  path: string;\n  /** Optional. Model file type/extension (for example `task`, `bin`, `litertlm`, or `gemini-nano`). If not provided, it is extracted from the path. Use `litertlm` on iOS only in SwiftPM integrations. */\n  modelType?: string;\n  /** Maximum number of tokens the model handles */\n  maxTokens?: number;\n  /** Number of tokens the model considers at each step */\n  topk?: number;\n  /** Amount of randomness in generation (0.0-1.0) */\n  temperature?: number;\n  /** Optional. Random seed for generation. */\n  randomSeed?: number;\n}\n"]}
+{"version":3,"file":"definitions.js","sourceRoot":"","sources":["../../src/definitions.ts"],"names":[],"mappings":"","sourcesContent":["/**\n * LLM Plugin interface for interacting with on-device language models\n */\nexport interface LLMPlugin {\n  /**\n   * Creates a new chat session\n   * @returns Promise with chat id and optional instructions\n   */\n  createChat(): Promise<{ id: string; instructions?: string }>;\n\n  /**\n   * Sends a message to the AI in a specific chat session\n   * @param options - The chat id and message to send\n   * @returns Promise that resolves when message is sent\n   */\n  sendMessage(options: { chatId: string; message: string }): Promise<void>;\n\n  /**\n   * Gets the readiness status of the LLM\n   * @returns Promise with readiness status string\n   */\n  getReadiness(): Promise<{ readiness: string }>;\n\n  /**\n   * Sets the model configuration\n   * - iOS: Use \"Apple Intelligence\" as path for the system model. Custom LiteRT-LM `.litertlm` models are supported on iOS only when this plugin is integrated through Swift Package Manager.\n   * - Android: Use \"Gemini Nano\" for the AICore system model on supported devices where it is already available. Prefer LiteRT-LM `.litertlm` bundles for custom models; legacy MediaPipe `.task` models are still supported\n   * - Web: Provide a web-ready model asset for `@mediapipe/tasks-genai` such as Gemma 4 `*-web.task`\n   * @param options - The model configuration\n   * @returns Promise that resolves when model is loaded\n   */\n  setModel(options: ModelOptions): Promise<void>;\n\n  /**\n   * Downloads a model from a URL and saves it to the appropriate location\n   * - iOS: Downloads to the app's documents directory\n   * - Android: Downloads to the app's files directory\n   * @param options - The download configuration\n   * @returns Promise with download result including the saved path\n   */\n  downloadModel(options: DownloadModelOptions): Promise<DownloadModelResult>;\n\n  /**\n   * Adds a listener for text received from AI\n   * @param eventName - Event name 'textFromAi'\n   * @param listenerFunc - Callback function for text events\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'textFromAi',\n    listenerFunc: (event: TextFromAiEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Adds a listener for AI completion events\n   * @param eventName - Event name 'aiFinished'\n   * @param listenerFunc - Callback function for finish events\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'aiFinished',\n    listenerFunc: (event: AiFinishedEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Adds a listener for generation failures that happen after streaming starts\n   * @param eventName - Event name 'generationError'\n   * @param listenerFunc - Callback function for generation errors\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'generationError',\n    listenerFunc: (event: GenerationErrorEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Adds a listener for model download progress events\n   * @param eventName - Event name 'downloadProgress'\n   * @param listenerFunc - Callback function for progress events\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'downloadProgress',\n    listenerFunc: (event: DownloadProgressEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Adds a listener for readiness status changes\n   * @param eventName - Event name 'readinessChange'\n   * @param listenerFunc - Callback function for readiness events\n   * @returns Promise with remove function to unsubscribe\n   */\n  addListener(\n    eventName: 'readinessChange',\n    listenerFunc: (event: ReadinessChangeEvent) => void,\n  ): Promise<{ remove: () => Promise<void> }>;\n\n  /**\n   * Get the native Capacitor plugin version.\n   *\n   * @returns Promise that resolves with the plugin version\n   * @throws Error if getting the version fails\n   * @since 1.0.0\n   * @example\n   * ```typescript\n   * const { version } = await CapgoLLM.getPluginVersion();\n   * console.log('Plugin version:', version);\n   * ```\n   */\n  getPluginVersion(): Promise<{ version: string }>;\n}\n\n/**\n * Event data for text received from AI\n */\nexport interface TextFromAiEvent {\n  /** The text content from AI - this is an incremental chunk, not the full text */\n  text: string;\n  /** The chat session ID */\n  chatId: string;\n  /** Whether this is a complete chunk (true) or partial streaming data (false) */\n  isChunk?: boolean;\n}\n\n/**\n * Event data for AI completion\n */\nexport interface AiFinishedEvent {\n  /** The chat session ID that finished */\n  chatId: string;\n}\n\n/**\n * Event data for generation failures\n * `chatId` is optional and may be omitted when the failure is not tied to a specific chat.\n */\nexport interface GenerationErrorEvent {\n  /** Optional. The chat session ID that failed, when available. */\n  chatId?: string;\n  /** Error message describing the failure */\n  error: string;\n}\n\n/**\n * Options for downloading a model\n * Only `url` is required. `companionUrl` and `filename` are optional.\n */\nexport interface DownloadModelOptions {\n  /** URL of the model file to download */\n  url: string;\n  /** Optional: URL of a companion file for legacy model formats */\n  companionUrl?: string;\n  /** Optional: Custom filename (defaults to filename from URL) */\n  filename?: string;\n}\n\n/**\n * Result of model download\n */\nexport interface DownloadModelResult {\n  /** Path where the model was saved */\n  path: string;\n  /** Path where the companion file was saved (if applicable) */\n  companionPath?: string;\n}\n\n/**\n * Event data for download progress\n */\nexport interface DownloadProgressEvent {\n  /** Percentage of download completed (0-100) */\n  progress: number;\n  /** Total bytes to download */\n  totalBytes?: number;\n  /** Bytes downloaded so far */\n  downloadedBytes?: number;\n}\n\n/**\n * Event data for readiness status changes\n */\nexport interface ReadinessChangeEvent {\n  /** The readiness status */\n  readiness: string;\n}\n\n/**\n * Model configuration options\n * Only `path` is required. All other properties are optional overrides.\n */\nexport interface ModelOptions {\n  /** Model path, \"Apple Intelligence\" for the Apple system model on iOS, or \"Gemini Nano\" for the Android AICore system model when already available on the device. On iOS, custom `.litertlm` models require the plugin to be integrated through Swift Package Manager. Gemma 4 examples use `.litertlm` on mobile and `*-web.task` on web. */\n  path: string;\n  /** Optional. Model file type/extension (for example `task`, `bin`, `litertlm`, or `gemini-nano`). If not provided, it is extracted from the path. Use `litertlm` on iOS only in SwiftPM integrations. */\n  modelType?: string;\n  /** Maximum number of tokens the model handles */\n  maxTokens?: number;\n  /** Number of tokens the model considers at each step */\n  topk?: number;\n  /** Amount of randomness in generation (0.0-1.0) */\n  temperature?: number;\n  /** Optional. Random seed for generation. */\n  randomSeed?: number;\n  /**\n   * Optional. LiteRT-LM engine backend for iOS (SwiftPM) and Android.\n   * Use `cpu` for stable long generations. When omitted, iOS prefers CPU then falls back to GPU; Android uses CPU.\n   * @since 8.2.0\n   */\n  backend?: 'gpu' | 'cpu';\n}\n"]}

package/ios/Sources/LLMPlugin/LLMPlugin.swift

@@ -17,7 +17,7 @@ import MediaPipeTasksGenAI
 
 @objc(LLMPlugin)
 public class LLMPlugin: CAPPlugin, CAPBridgedPlugin {
-    private let pluginVersion: String = "8.1.0"
+    private let pluginVersion: String = "8.1.1"
     private let liteRtTopP: Float = 0.95
 
     public let identifier = "LLMPlugin"
@@ -80,6 +80,7 @@ public class LLMPlugin: CAPPlugin, CAPBridgedPlugin {
         let temperature = call.getFloat("temperature") ?? 0.8
         let randomSeed = call.getInt("randomSeed") ?? 0
         let modelType = call.getString("modelType")
+        let backend = call.getString("backend")
 
         modelPath = path
         isReady = false
@@ -112,7 +113,11 @@ public class LLMPlugin: CAPPlugin, CAPBridgedPlugin {
             Task {
                 do {
                     let modelURL = try resolveModelURL(path: path, modelType: modelType)
-                    let engine = try await initializeLiteRtEngine(modelURL: modelURL, maxTokens: maxTokens)
+                    let engine = try await initializeLiteRtEngine(
+                        modelURL: modelURL,
+                        maxTokens: maxTokens,
+                        backend: backend
+                    )
                     liteRtEngine = engine
                     liteRtTopk = topk
                     liteRtTemperature = temperature
@@ -528,7 +533,11 @@ private extension LLMPlugin {
 
 #if canImport(LiteRTLM)
 private extension LLMPlugin {
-    func initializeLiteRtEngine(modelURL: URL, maxTokens: Int) async throws -> LiteRTLM.Engine {
+    func initializeLiteRtEngine(
+        modelURL: URL,
+        maxTokens: Int,
+        backend: String?
+    ) async throws -> LiteRTLM.Engine {
         guard let cacheDirectory = FileManager.default.urls(for: .cachesDirectory, in: .userDomainMask).first else {
             throw NSError(
                 domain: "CapgoLLM",
@@ -537,20 +546,25 @@ private extension LLMPlugin {
             )
         }
 
-        let configs = [
+        let cacheDir = cacheDirectory.path
+        let makeConfig = { (selectedBackend: LiteRTLM.Backend) throws -> LiteRTLM.EngineConfig in
             try LiteRTLM.EngineConfig(
                 modelPath: modelURL.path,
-                backend: .gpu,
+                backend: selectedBackend,
                 maxNumTokens: maxTokens,
-                cacheDir: cacheDirectory.path
-            ),
-            try LiteRTLM.EngineConfig(
-                modelPath: modelURL.path,
-                backend: .cpu(),
-                maxNumTokens: maxTokens,
-                cacheDir: cacheDirectory.path
+                cacheDir: cacheDir
             )
-        ]
+        }
+
+        let configs: [LiteRTLM.EngineConfig]
+        switch backend?.lowercased() {
+        case "cpu":
+            configs = [try makeConfig(.cpu())]
+        case "gpu":
+            configs = [try makeConfig(.gpu), try makeConfig(.cpu())]
+        default:
+            configs = [try makeConfig(.cpu()), try makeConfig(.gpu)]
+        }
 
         var lastError: Error?
         for config in configs {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@capgo/capacitor-llm",
-  "version": "8.1.0",
+  "version": "8.1.1",
   "description": "Adds support for LLM locally runned for Capacitor",
   "main": "dist/plugin.cjs.js",
   "module": "dist/esm/index.js",