@dr33m/react-native-litert-lm 0.5.3 → 0.5.4

package/android/src/main/java/com/margelo/nitro/dev/litert/litertlm/HybridLiteRTLM.kt

@@ -58,6 +58,10 @@ class HybridLiteRTLM : HybridLiteRTLMSpec() {
         /** Cached result of OpenCL availability probe (null = not yet checked). */
         @Volatile
         private var openCLAvailable: Boolean? = null
+
+        /** Cached result of NPU/QNN availability probe (null = not yet checked). */
+        @Volatile
+        private var npuAvailable: Boolean? = null
         
         /**
          * Initialize the native library.
@@ -197,6 +201,41 @@ class HybridLiteRTLM : HybridLiteRTLMSpec() {
                         }
                     }
 
+                    // NPU hardware check: probe for QNN HTP runtime libraries.
+                    // LiteRT-LM's NPU delegate requires Qualcomm QNN HTP (Hexagon Tensor
+                    // Processor) runtime. Without it, Engine() crashes with SIGSEGV
+                    // that Kotlin's try/catch cannot intercept.
+                    // Unlike GPU (which throws a catchable Java exception on failure),
+                    // NPU failure is a native crash — so we MUST detect before attempting.
+                    if (npuAvailable == null) {
+                        val hasNpu = run {
+                            // Check for QNN HTP libraries in system vendor paths.
+                            // These are only present on devices with Qualcomm NPU support.
+                            val qnnPaths = arrayOf(
+                                "/vendor/lib64/libQnnHtp.so",
+                                "/vendor/lib/libQnnHtp.so",
+                                "/system/vendor/lib64/libQnnHtp.so",
+                                "/system/lib64/libQnnHtp.so",
+                                "/vendor/lib64/libQnnSystem.so",
+                                "/vendor/lib/libQnnSystem.so"
+                            )
+                            var found = false
+                            for (path in qnnPaths) {
+                                if (java.io.File(path).exists()) {
+                                    found = true
+                                    break
+                                }
+                            }
+                            found
+                        }
+                        npuAvailable = hasNpu
+                        if (!hasNpu) {
+                            Log.w(TAG, "QNN HTP libraries not found — NPU backend unavailable")
+                        } else {
+                            Log.i(TAG, "QNN HTP libraries found — NPU backend may be available")
+                        }
+                    }
+
                     // Detect multimodal support. Check config.multimodal flag first, then fall back to filename sniffing.
                     // Only Gemma 3n bundles vision/audio executors; Gemma 4 E2B is text-only.
                     // Passing vision/audio backends to a text-only model causes
@@ -259,12 +298,13 @@ class HybridLiteRTLM : HybridLiteRTLMSpec() {
                             }
                         }
                         Backend.NPU -> {
+                            val hasNpu = npuAvailable ?: false
                             val nativeLibDir = LiteRTLMInitProvider.applicationContext?.applicationInfo?.nativeLibraryDir
-                            Log.i(TAG, "NPU backend requested - nativeLibraryDir=$nativeLibDir")
-                            if (nativeLibDir != null) {
+                            Log.i(TAG, "NPU backend requested - available=$hasNpu, nativeLibraryDir=$nativeLibDir")
+                            if (hasNpu && nativeLibDir != null) {
                                 com.google.ai.edge.litertlm.Backend.NPU(nativeLibraryDir = nativeLibDir)
                             } else {
-                                Log.w(TAG, "NPU requested but nativeLibraryDir unavailable — using CPU directly")
+                                Log.w(TAG, "NPU requested but hardware unavailable — using CPU directly")
                                 backend = Backend.CPU
                                 com.google.ai.edge.litertlm.Backend.CPU()
                             }
@@ -662,10 +702,10 @@ class HybridLiteRTLM : HybridLiteRTLMSpec() {
             tool(apiTool)
         }
 
-        // Create conversation with explicit SamplerConfig (required by Gallery pattern).
-        // GPU backend may fail silently without proper sampler params.
+        // Create conversation config. NPU backend does not support SamplerConfig
+        // (matching Gallery app pattern — setting sampler params on NPU causes crashes).
         val convConfig = ConversationConfig(
-            samplerConfig = SamplerConfig(
+            samplerConfig = if (backend == Backend.NPU) null else SamplerConfig(
                 topK = topK,
                 topP = topP.toDouble(),
                 temperature = temperature.toDouble(),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dr33m/react-native-litert-lm",
-  "version": "0.5.3",
+  "version": "0.5.4",
   "litertLm": {
     "version": "0.12.0",
     "androidMavenVersion": "0.12.0",