Fix bugs across Kotlin, Flutter, and React Native SDKs

Siddhesh2377 · claude · Siddhesh2377 · commit c167ea724afe · 2026-03-18T23:40:33.000+05:30
Genie NPU fixes (all 3 SDKs): - downloadUrl() now accepts quant parameter (default w4a16) for w8a16 models - Fixed model slug mismatch (llama-v3.2 → llama3.2) - Added missing models: SEA-LION v3.5 8B, Qwen 2.5 7B (w8a16) Kotlin SDK: - SecureStorageFactory.create() race condition → @synchronized - SDKLogger destination iteration thread safety → .toList() snapshot - isNativeInitialized() now calls RunAnywhereBridge.racIsInitialized() - Removed runBlocking ANR risk → suspend chain with Mutex React Native SDK: - Memory leak: added rac_llm_result_free() in generate() - cancelDownload() now actually stops RNFS download via jobId tracking - CMakeLists RAC_INCLUDE_DIR fixed to jniLibs/include/ (3 files) - PlatformAdapterBridge.getChipName() uses Build.SOC_MODEL (API 31+) - extractIntValue/extractFloatValue crash guard (try/catch) - transcribeFile returns JSON matching transcribe() + fixed JSON escaping - WAV parser: memcpy instead of reinterpret_cast (ARM safe) - lastError_ protected with errorMutex_ - deleteModel handles all file extensions - getModelInfo handles empty object as null - Fixed VLM cross-package import error Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
diff --git a/examples/android/RunAnywhereAI/app/src/main/java/com/runanywhere/runanywhereai/data/ModelList.kt b/examples/android/RunAnywhereAI/app/src/main/java/com/runanywhere/runanywhereai/data/ModelList.kt
@@ -137,6 +137,7 @@ object ModelList {
         val name: String,
         val memoryRequirement: Long,
         val supportedChips: Set<NPUChip>,
+        val quant: String = "w4a16",
     )
 
     private val genieModelDefs = listOf(
@@ -147,11 +148,24 @@ object ModelList {
             supportedChips = setOf(NPUChip.SNAPDRAGON_8_ELITE_GEN5),
         ),
         GenieModelDef(
-            slug = "llama-v3.2-1b-instruct",
+            slug = "llama3.2-1b-instruct",
             name = "Llama 3.2 1B Instruct",
             memoryRequirement = 1_200_000_000,
             supportedChips = setOf(NPUChip.SNAPDRAGON_8_ELITE, NPUChip.SNAPDRAGON_8_ELITE_GEN5),
         ),
+        GenieModelDef(
+            slug = "sea-lion3.5-8b-instruct",
+            name = "SEA-LION v3.5 8B Instruct",
+            memoryRequirement = 4_800_000_000,
+            supportedChips = setOf(NPUChip.SNAPDRAGON_8_ELITE, NPUChip.SNAPDRAGON_8_ELITE_GEN5),
+        ),
+        GenieModelDef(
+            slug = "qwen2.5-7b-instruct",
+            name = "Qwen 2.5 7B Instruct",
+            memoryRequirement = 4_200_000_000,
+            supportedChips = setOf(NPUChip.SNAPDRAGON_8_ELITE),
+            quant = "w8a16",
+        ),
     )
 
     private fun genieModels(): List<AppModel> {
@@ -162,7 +176,7 @@ object ModelList {
                 AppModel(
                     id = "${def.slug}-npu-${chip.identifier}",
                     name = "${def.name} (NPU - ${chip.displayName})",
-                    url = chip.downloadUrl(def.slug),
+                    url = chip.downloadUrl(def.slug, def.quant),
                     framework = InferenceFramework.GENIE,
                     category = ModelCategory.LANGUAGE,
                     memoryRequirement = def.memoryRequirement,
diff --git a/examples/flutter/RunAnywhereAI/lib/app/runanywhere_ai_app.dart b/examples/flutter/RunAnywhereAI/lib/app/runanywhere_ai_app.dart
@@ -215,17 +215,20 @@ class _RunAnywhereAIAppState extends State<RunAnywhereAIApp> {
         // Models with per-chip availability
         const genieModels = [
           // Qwen3 4B — Gen 5 only
-          (slug: 'qwen3-4b', name: 'Qwen3 4B', mem: 2800000000, chips: {NPUChip.snapdragon8EliteGen5}),
+          (slug: 'qwen3-4b', name: 'Qwen3 4B', mem: 2800000000, quant: 'w4a16', chips: {NPUChip.snapdragon8EliteGen5}),
           // Llama 3.2 1B Instruct — both chips
-          // TODO: Remove snapdragon7sGen3Test before merging
-          (slug: 'llama-v3.2-1b-instruct', name: 'Llama 3.2 1B Instruct', mem: 1200000000, chips: {NPUChip.snapdragon8Elite, NPUChip.snapdragon8EliteGen5, NPUChip.snapdragon7sGen3Test}),
+          (slug: 'llama3.2-1b-instruct', name: 'Llama 3.2 1B Instruct', mem: 1200000000, quant: 'w4a16', chips: {NPUChip.snapdragon8Elite, NPUChip.snapdragon8EliteGen5}),
+          // SEA-LION v3.5 8B Instruct — both chips
+          (slug: 'sea-lion3.5-8b-instruct', name: 'SEA-LION v3.5 8B Instruct', mem: 4800000000, quant: 'w4a16', chips: {NPUChip.snapdragon8Elite, NPUChip.snapdragon8EliteGen5}),
+          // Qwen 2.5 7B Instruct — 8elite only, w8a16 quant
+          (slug: 'qwen2.5-7b-instruct', name: 'Qwen 2.5 7B Instruct', mem: 4200000000, quant: 'w8a16', chips: {NPUChip.snapdragon8Elite}),
         ];
         for (final m in genieModels) {
           if (m.chips.contains(chip)) {
             Genie.addModel(
               id: '${m.slug}-npu-${chip.identifier}',
               name: '${m.name} (NPU - ${chip.displayName})',
-              url: chip.downloadUrl(m.slug),
+              url: chip.downloadUrl(m.slug, quant: m.quant),
               memoryRequirement: m.mem,
             );
           }
diff --git a/examples/react-native/RunAnywhereAI/App.tsx b/examples/react-native/RunAnywhereAI/App.tsx
@@ -256,11 +256,16 @@ async function registerModulesAndModels(): Promise<void> {
         name: string;
         mem: number;
         chips: string[];
+        quant?: string;
       }> = [
         // Qwen3 4B — Gen 5 only
         { slug: 'qwen3-4b', name: 'Qwen3 4B', mem: 2_800_000_000, chips: ['8elite-gen5'] },
         // Llama 3.2 1B Instruct — both chips
-        { slug: 'llama-v3.2-1b-instruct', name: 'Llama 3.2 1B Instruct', mem: 1_200_000_000, chips: ['8elite', '8elite-gen5'] },
+        { slug: 'llama3.2-1b-instruct', name: 'Llama 3.2 1B Instruct', mem: 1_200_000_000, chips: ['8elite', '8elite-gen5'] },
+        // SEA-LION v3.5 8B Instruct — both chips
+        { slug: 'sea-lion3.5-8b-instruct', name: 'SEA-LION v3.5 8B Instruct', mem: 4_800_000_000, chips: ['8elite', '8elite-gen5'] },
+        // Qwen 2.5 7B Instruct — 8elite only, w8a16 quant
+        { slug: 'qwen2.5-7b-instruct', name: 'Qwen 2.5 7B Instruct', mem: 4_200_000_000, chips: ['8elite'], quant: 'w8a16' },
       ];
 
       const registrations = genieModels
@@ -269,7 +274,7 @@ async function registerModulesAndModels(): Promise<void> {
           RunAnywhere.registerModel({
             id: `${m.slug}-npu-${chip.identifier}`,
             name: `${m.name} (NPU - ${chip.displayName})`,
-            url: getNPUDownloadUrl(chip, m.slug),
+            url: getNPUDownloadUrl(chip, m.slug, m.quant),
             framework: LLMFramework.Genie,
             memoryRequirement: m.mem,
           }),
diff --git a/sdk/runanywhere-flutter/packages/runanywhere/lib/core/types/npu_chip.dart b/sdk/runanywhere-flutter/packages/runanywhere/lib/core/types/npu_chip.dart
@@ -29,8 +29,9 @@ enum NPUChip {
   /// Build a HuggingFace download URL for this chip.
   /// [modelSlug] is the model slug (e.g. "qwen3-4b") → produces
   ///   "qwen3-4b-genie-w4a16-8elite-gen5.tar.gz"
-  String downloadUrl(String modelSlug) =>
-      '$baseUrl$modelSlug-genie-w4a16-$npuSuffix.tar.gz';
+  /// [quant] is the quantization format (e.g. "w4a16", "w8a16"). Defaults to "w4a16".
+  String downloadUrl(String modelSlug, {String quant = 'w4a16'}) =>
+      '$baseUrl$modelSlug-genie-$quant-$npuSuffix.tar.gz';
 
   /// Match an NPU chip from a SoC model string (e.g. "SM8750").
   /// Returns null if the SoC is not a supported NPU chipset.
diff --git a/sdk/runanywhere-kotlin/src/androidMain/kotlin/com/runanywhere/sdk/security/SecureStorage.kt b/sdk/runanywhere-kotlin/src/androidMain/kotlin/com/runanywhere/sdk/security/SecureStorage.kt
@@ -39,6 +39,7 @@ class AndroidSecureStorage private constructor(
         /**
          * Create secure storage instance for Android
          */
+        @Synchronized
         fun create(identifier: String): AndroidSecureStorage {
             val appContext =
                 context
diff --git a/sdk/runanywhere-kotlin/src/commonMain/kotlin/com/runanywhere/sdk/core/types/NPUChip.kt b/sdk/runanywhere-kotlin/src/commonMain/kotlin/com/runanywhere/sdk/core/types/NPUChip.kt
@@ -27,9 +27,10 @@ enum class NPUChip(
      * Build a HuggingFace download URL for this chip.
      * @param modelSlug Model slug (e.g. "qwen3-4b") → produces
      *   "qwen3-4b-genie-w4a16-8elite-gen5.tar.gz"
+     * @param quant Quantization format (e.g. "w4a16", "w8a16"). Defaults to "w4a16".
      */
-    fun downloadUrl(modelSlug: String): String =
-        "${BASE_URL}${modelSlug}-genie-w4a16-${npuSuffix}.tar.gz"
+    fun downloadUrl(modelSlug: String, quant: String = "w4a16"): String =
+        "${BASE_URL}${modelSlug}-genie-${quant}-${npuSuffix}.tar.gz"
 
     companion object {
         /** Base URL for NPU model downloads on HuggingFace. */
diff --git a/sdk/runanywhere-kotlin/src/commonMain/kotlin/com/runanywhere/sdk/foundation/SDKLogger.kt b/sdk/runanywhere-kotlin/src/commonMain/kotlin/com/runanywhere/sdk/foundation/SDKLogger.kt
@@ -332,8 +332,8 @@ object Logging {
         // Forward to runanywhere-commons bridge if set
         commonsLogBridge?.invoke(entry)
 
-        // Write to all registered destinations
-        for (destination in _destinations) {
+        // Write to all registered destinations (snapshot to avoid concurrent modification)
+        for (destination in _destinations.toList()) {
             if (destination.isAvailable) {
                 destination.write(entry)
             }
@@ -358,6 +358,7 @@ object Logging {
     /**
      * Add a log destination (non-suspending version).
      */
+    @Synchronized
     fun addDestinationSync(destination: LogDestination) {
         if (_destinations.none { it.identifier == destination.identifier }) {
             _destinations.add(destination)
diff --git a/sdk/runanywhere-kotlin/src/commonMain/kotlin/com/runanywhere/sdk/public/RunAnywhere.kt b/sdk/runanywhere-kotlin/src/commonMain/kotlin/com/runanywhere/sdk/public/RunAnywhere.kt
@@ -14,6 +14,8 @@ import com.runanywhere.sdk.foundation.LogLevel
 import com.runanywhere.sdk.foundation.SDKLogger
 import com.runanywhere.sdk.public.events.EventBus
 import com.runanywhere.sdk.utils.SDKConstants
+import kotlinx.coroutines.sync.Mutex
+import kotlinx.coroutines.sync.withLock
 
 // ═══════════════════════════════════════════════════════════════════════════
 // SDK INITIALIZATION FLOW (Two-Phase Pattern)
@@ -91,6 +93,7 @@ object RunAnywhere {
     private var _areServicesReady: Boolean = false
 
     private val lock = Any()
+    private val servicesMutex = Mutex()
 
     // ═══════════════════════════════════════════════════════════════════════════
     // MARK: - Public Properties
@@ -240,7 +243,7 @@ object RunAnywhere {
             return
         }
 
-        synchronized(lock) {
+        servicesMutex.withLock {
             if (_areServicesReady) {
                 return
             }
@@ -334,7 +337,7 @@ object RunAnywhere {
      * Initialize CppBridge services (Phase 2)
      * Implementation is in jvmAndroidMain via expect/actual
      */
-    private fun initializeCppBridgeServices() {
+    private suspend fun initializeCppBridgeServices() {
         logger.debug("CppBridge services initialization requested")
         initializePlatformBridgeServices()
     }
@@ -367,7 +370,7 @@ internal expect fun initializePlatformBridge(environment: SDKEnvironment, apiKey
  * Initialize platform-specific bridge services (Phase 2).
  * On JVM/Android, this calls CppBridge.initializeServices().
  */
-internal expect fun initializePlatformBridgeServices()
+internal expect suspend fun initializePlatformBridgeServices()
 
 /**
  * Shutdown platform-specific bridge.
diff --git a/sdk/runanywhere-kotlin/src/jvmAndroidMain/kotlin/com/runanywhere/sdk/foundation/bridge/CppBridge.kt b/sdk/runanywhere-kotlin/src/jvmAndroidMain/kotlin/com/runanywhere/sdk/foundation/bridge/CppBridge.kt
@@ -558,8 +558,12 @@ object CppBridge {
      * @return true if rac_is_initialized() returns true
      */
     fun isNativeInitialized(): Boolean {
-        // TODO: Call rac_is_initialized()
-        return _isInitialized
+        if (!_isInitialized || !isNativeLibraryLoaded) return false
+        return try {
+            RunAnywhereBridge.racIsInitialized()
+        } catch (_: Exception) {
+            _isInitialized
+        }
     }
 
     // =============================================================================
diff --git a/sdk/runanywhere-kotlin/src/jvmAndroidMain/kotlin/com/runanywhere/sdk/public/PlatformBridge.kt b/sdk/runanywhere-kotlin/src/jvmAndroidMain/kotlin/com/runanywhere/sdk/public/PlatformBridge.kt
@@ -11,7 +11,6 @@ package com.runanywhere.sdk.public
 import com.runanywhere.sdk.foundation.SDKLogger
 import com.runanywhere.sdk.foundation.bridge.CppBridge
 import com.runanywhere.sdk.foundation.bridge.extensions.CppBridgeTelemetry
-import kotlinx.coroutines.runBlocking
 
 private const val TAG = "PlatformBridge"
 private val logger = SDKLogger(TAG)
@@ -49,15 +48,9 @@ internal actual fun initializePlatformBridge(environment: SDKEnvironment, apiKey
  * Initialize CppBridge services (Phase 2).
  * This includes model assignment, platform services, and device registration.
  */
-internal actual fun initializePlatformBridgeServices() {
+internal actual suspend fun initializePlatformBridgeServices() {
     logger.info("Initializing CppBridge services...")
-
-    // Use runBlocking to call the suspend function
-    // This is safe because services initialization is typically called once
-    runBlocking {
-        CppBridge.initializeServices()
-    }
-
+    CppBridge.initializeServices()
     logger.info("CppBridge services initialization complete")
 }
 
diff --git a/sdk/runanywhere-react-native/packages/core/android/CMakeLists.txt b/sdk/runanywhere-react-native/packages/core/android/CMakeLists.txt
@@ -27,8 +27,8 @@ set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -Wl,-z,max-page-size
 # Path to pre-built native libraries (downloaded from runanywhere-sdks)
 set(JNILIB_DIR ${CMAKE_SOURCE_DIR}/src/main/jniLibs/${ANDROID_ABI})
 
-# Path to RAC headers (downloaded with native libraries)
-set(RAC_INCLUDE_DIR ${CMAKE_SOURCE_DIR}/src/main/include)
+# Path to RAC headers (downloaded with native libraries into jniLibs)
+set(RAC_INCLUDE_DIR ${CMAKE_SOURCE_DIR}/src/main/jniLibs/include)
 
 # =============================================================================
 # RACommons - Core SDK functionality (rac_* API)
diff --git a/sdk/runanywhere-react-native/packages/core/android/src/main/java/com/margelo/nitro/runanywhere/PlatformAdapterBridge.kt b/sdk/runanywhere-react-native/packages/core/android/src/main/java/com/margelo/nitro/runanywhere/PlatformAdapterBridge.kt
@@ -379,10 +379,16 @@ object PlatformAdapterBridge {
     }
 
     /**
-     * Get chip name (e.g., "Tensor G3")
+     * Get chip name (e.g., "SM8750")
      */
     @JvmStatic
     fun getChipName(): String {
+        if (android.os.Build.VERSION.SDK_INT >= android.os.Build.VERSION_CODES.S) {
+            val socModel = android.os.Build.SOC_MODEL
+            if (!socModel.isNullOrEmpty() && socModel != "unknown") {
+                return socModel
+            }
+        }
         return android.os.Build.HARDWARE
     }
 
diff --git a/sdk/runanywhere-react-native/packages/core/cpp/HybridRunAnywhereCore.cpp b/sdk/runanywhere-react-native/packages/core/cpp/HybridRunAnywhereCore.cpp
@@ -1310,7 +1310,10 @@ std::shared_ptr<Promise<std::string>> HybridRunAnywhereCore::httpGet(
 // ============================================================================
 
 std::shared_ptr<Promise<std::string>> HybridRunAnywhereCore::getLastError() {
-    return Promise<std::string>::async([this]() { return lastError_; });
+    return Promise<std::string>::async([this]() {
+        std::lock_guard<std::mutex> lock(errorMutex_);
+        return lastError_;
+    });
 }
 
 // Forward declaration for platform-specific archive extraction
@@ -1435,7 +1438,10 @@ std::shared_ptr<Promise<double>> HybridRunAnywhereCore::getMemoryUsage() {
 // ============================================================================
 
 void HybridRunAnywhereCore::setLastError(const std::string& error) {
-    lastError_ = error;
+    {
+        std::lock_guard<std::mutex> lock(errorMutex_);
+        lastError_ = error;
+    }
     LOGE("%s", error.c_str());
 }
 
@@ -1552,12 +1558,15 @@ std::shared_ptr<Promise<std::string>> HybridRunAnywhereCore::generate(
 
         std::string text = llmResult.text ? llmResult.text : "";
         int tokensUsed = llmResult.completion_tokens;
+        double latencyMs = llmResult.total_time_ms;
+
+        rac_llm_result_free(&llmResult);
 
         return buildJsonObject({
             {"text", jsonString(text)},
             {"tokensUsed", std::to_string(tokensUsed)},
             {"modelUsed", jsonString("llm")},
-            {"latencyMs", std::to_string(llmResult.total_time_ms)}
+            {"latencyMs", std::to_string(latencyMs)}
         });
     });
 }
@@ -1984,11 +1993,12 @@ std::shared_ptr<Promise<std::string>> HybridRunAnywhereCore::transcribeFile(
             while (pos + 8 < dataSize) {
                 char chunkId[5] = {0};
                 memcpy(chunkId, &data[pos], 4);
-                uint32_t chunkSize = *reinterpret_cast<const uint32_t*>(&data[pos + 4]);
+                uint32_t chunkSize;
+                memcpy(&chunkSize, &data[pos + 4], sizeof(chunkSize));
 
                 if (strcmp(chunkId, "fmt ") == 0) {
                     if (pos + 8 + chunkSize <= dataSize && chunkSize >= 16) {
-                        sampleRate = *reinterpret_cast<const int32_t*>(&data[pos + 12]);
+                        memcpy(&sampleRate, &data[pos + 12], sizeof(sampleRate));
                         if (sampleRate <= 0 || sampleRate > 48000) sampleRate = 16000;
                         LOGI("WAV sample rate: %d Hz", sampleRate);
                     }
@@ -2042,11 +2052,11 @@ std::shared_ptr<Promise<std::string>> HybridRunAnywhereCore::transcribeFile(
 
             rac_stt_result_free(&result);
             LOGI("Transcription result: %s", transcribedText.c_str());
-            return transcribedText;
+            return "{\"text\":" + jsonString(transcribedText) + ",\"confidence\":0}";
         } catch (const std::exception& e) {
             std::string msg = e.what();
             LOGI("TranscribeFile exception: %s", msg.c_str());
-            return "{\"error\":\"" + msg + "\"}";
+            return "{\"error\":" + jsonString(msg) + "}";
         } catch (...) {
             return "{\"error\":\"Transcription failed (unknown error)\"}";
         }
diff --git a/sdk/runanywhere-react-native/packages/core/cpp/HybridRunAnywhereCore.hpp b/sdk/runanywhere-react-native/packages/core/cpp/HybridRunAnywhereCore.hpp
@@ -292,6 +292,7 @@ class HybridRunAnywhereCore : public HybridRunAnywhereCoreSpec {
 private:
   // Thread safety
   std::mutex initMutex_;
+  std::mutex errorMutex_;
 
   // State tracking
   std::string lastError_;
diff --git a/sdk/runanywhere-react-native/packages/core/src/Public/Extensions/RunAnywhere+Models.ts b/sdk/runanywhere-react-native/packages/core/src/Public/Extensions/RunAnywhere+Models.ts
@@ -85,7 +85,10 @@ export async function getModelInfo(modelId: string): Promise<ModelInfo | null> {
   const infoJson = await native.getModelInfo(modelId);
   try {
     const result = JSON.parse(infoJson);
-    return result === 'null' ? null : result;
+    if (result === null || result === 'null' || (typeof result === 'object' && Object.keys(result).length === 0)) {
+      return null;
+    }
+    return result;
   } catch {
     return null;
   }
@@ -508,8 +511,10 @@ export async function downloadModel(
  */
 export async function cancelDownload(modelId: string): Promise<boolean> {
   if (activeDownloads.has(modelId)) {
+    // Stop the native RNFS download job
+    FileSystem.cancelDownload(modelId);
     activeDownloads.delete(modelId);
-    logger.info(`Marked download as cancelled: ${modelId}`);
+    logger.info(`Cancelled download: ${modelId}`);
     return true;
   }
   return false;
@@ -521,9 +526,19 @@ export async function cancelDownload(modelId: string): Promise<boolean> {
 export async function deleteModel(modelId: string): Promise<boolean> {
   try {
     const modelInfo = await ModelRegistry.getModel(modelId);
-    const extension = modelInfo?.downloadURL?.includes('.gguf')
-      ? '.gguf'
-      : '';
+    const url = modelInfo?.downloadURL ?? '';
+    let extension = '';
+    if (url.includes('.gguf')) {
+      extension = '.gguf';
+    } else if (url.includes('.onnx')) {
+      extension = '.onnx';
+    } else if (url.includes('.tar.bz2')) {
+      extension = '.tar.bz2';
+    } else if (url.includes('.tar.gz')) {
+      extension = '.tar.gz';
+    } else if (url.includes('.zip')) {
+      extension = '.zip';
+    }
     const fileName = `${modelId}${extension}`;
 
     // Delete using FileSystem service
diff --git a/sdk/runanywhere-react-native/packages/core/src/Public/Extensions/RunAnywhere+VLM.ts b/sdk/runanywhere-react-native/packages/core/src/Public/Extensions/RunAnywhere+VLM.ts
diff --git a/sdk/runanywhere-react-native/packages/core/src/services/FileSystem.ts b/sdk/runanywhere-react-native/packages/core/src/services/FileSystem.ts
diff --git a/sdk/runanywhere-react-native/packages/core/src/types/NPUChip.ts b/sdk/runanywhere-react-native/packages/core/src/types/NPUChip.ts
diff --git a/sdk/runanywhere-react-native/packages/llamacpp/android/CMakeLists.txt b/sdk/runanywhere-react-native/packages/llamacpp/android/CMakeLists.txt
diff --git a/sdk/runanywhere-react-native/packages/llamacpp/cpp/HybridRunAnywhereLlama.cpp b/sdk/runanywhere-react-native/packages/llamacpp/cpp/HybridRunAnywhereLlama.cpp
diff --git a/sdk/runanywhere-react-native/packages/onnx/android/CMakeLists.txt b/sdk/runanywhere-react-native/packages/onnx/android/CMakeLists.txt
