@dr33m/react-native-litert-lm 0.5.5 → 0.5.6

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

@@ -575,7 +575,10 @@ class HybridLiteRTLM : HybridLiteRTLMSpec() {
         return lastStats
     }
 
-    override fun getMemoryUsage(): MemoryUsage {
+    /** Shared by the public [getMemoryUsage] override and the pre-flight guard
+     * in [execute] — both need the same real, OS-level reading, not an
+     * estimate. */
+    private fun currentMemoryUsage(): MemoryUsage {
         // Native heap: allocated bytes from Debug APIs (most accurate for native allocations)
         val nativeHeapBytes = Debug.getNativeHeapAllocatedSize().toDouble()
 
@@ -619,6 +622,8 @@ class HybridLiteRTLM : HybridLiteRTLMSpec() {
         )
     }
 
+    override fun getMemoryUsage(): MemoryUsage = currentMemoryUsage()
+
     override fun checkModelCapabilities(modelPath: String): ModelCapabilities {
         var supportsSpeculativeDecoding = false
         try {
@@ -784,6 +789,16 @@ class HybridLiteRTLM : HybridLiteRTLMSpec() {
 
         return Promise.parallel {
             ensureLoaded()
+
+            // Refuse rather than risk it: growing the KV-cache under real memory
+            // pressure can fail at the native engine level in a way that never
+            // surfaces as a catchable Kotlin exception — it takes the whole
+            // process down. A plain RuntimeException here is a normal rejected
+            // Promise for callers, same contract as any other execute() failure.
+            if (currentMemoryUsage().isLowMemory) {
+                throw RuntimeException("LiteRTLM: Device memory is critically low; refusing to continue generation.")
+            }
+
             // Clear any previous tool calls before new inference
             pendingToolCalls.clear()
 

package/ios/HybridLiteRTLM+Execute.swift

@@ -80,6 +80,18 @@ extension HybridLiteRTLM {
                 return
             }
 
+            // Refuse rather than risk it: growing the KV-cache under real memory
+            // pressure can fail inside the underlying C engine in a way that
+            // never surfaces as a catchable Swift error — it takes the whole
+            // process down (Jetsam or a hard native failure). An NSError here
+            // is a normal rejected Promise for callers, same contract as any
+            // other execute() failure.
+            if Self.currentMemoryUsage().isLowMemory {
+                promise.reject(withError: NSError(domain: "LiteRTLM", code: 507,
+                    userInfo: [NSLocalizedDescriptionKey: "LiteRTLM: Device memory is critically low; refusing to continue generation."]))
+                return
+            }
+
             let payload: (json: String, tempFiles: [String])
             do { payload = try self.buildExecutePayload(preprocessed) }
             catch { promise.reject(withError: error); return }

package/ios/HybridLiteRTLM.swift

@@ -111,9 +111,18 @@ public class HybridLiteRTLM: HybridLiteRTLMSpec_base, HybridLiteRTLMSpec_protoco
     }
     
     public func getMemoryUsage() throws -> MemoryUsage {
+        return Self.currentMemoryUsage()
+    }
+
+    /// Shared by the public `getMemoryUsage()` override and the pre-flight
+    /// guard in `HybridLiteRTLM+Execute.swift` — both need the same real,
+    /// OS-level reading, not an estimate. `static` since it needs no
+    /// instance state, which also makes it trivially callable from the
+    /// extension file.
+    static func currentMemoryUsage() -> MemoryUsage {
         var residentBytes: Double = 0.0
         var nativeHeapBytes: Double = 0.0
-        
+
         // Retrieve process resident set size (RSS) via Mach basic task info
         var info = mach_task_basic_info()
         var count = mach_msg_type_number_t(MemoryLayout<mach_task_basic_info>.size / MemoryLayout<integer_t>.size)
@@ -122,18 +131,18 @@ public class HybridLiteRTLM: HybridLiteRTLMSpec_base, HybridLiteRTLMSpec_protoco
                 task_info(mach_task_self_, task_flavor_t(MACH_TASK_BASIC_INFO), $0, &count)
             }
         }
-        
+
         if kerr == KERN_SUCCESS {
             residentBytes = Double(info.resident_size)
             nativeHeapBytes = Double(info.resident_size)
         }
-        
+
         // os_proc_available_memory reports actual headroom available before Jetsam termination (iOS 13+)
         let availableBytes = Double(os_proc_available_memory())
-        
+
         // Flag memory warning at ~200MB remaining headroom
         let isLowMemory = availableBytes < 200.0 * 1024.0 * 1024.0
-        
+
         return MemoryUsage(
             nativeHeapBytes: nativeHeapBytes,
             residentBytes: residentBytes,

package/nitrogen/generated/android/kotlin/com/margelo/nitro/dev/litert/litertlm/Func_void_double.kt

@@ -10,7 +10,6 @@ package com.margelo.nitro.dev.litert.litertlm
 import androidx.annotation.Keep
 import com.facebook.jni.HybridData
 import com.facebook.proguard.annotations.DoNotStrip
-import dalvik.annotation.optimization.FastNative
 
 
 /**
@@ -59,7 +58,6 @@ class Func_void_double_cxx: Func_void_double {
   override fun invoke(progress: Double): Unit
     = invoke_cxx(progress)
 
-  @FastNative
   private external fun invoke_cxx(progress: Double): Unit
 }
 

package/nitrogen/generated/android/kotlin/com/margelo/nitro/dev/litert/litertlm/Func_void_std__string_bool.kt

@@ -10,7 +10,6 @@ package com.margelo.nitro.dev.litert.litertlm
 import androidx.annotation.Keep
 import com.facebook.jni.HybridData
 import com.facebook.proguard.annotations.DoNotStrip
-import dalvik.annotation.optimization.FastNative
 
 
 /**
@@ -59,7 +58,6 @@ class Func_void_std__string_bool_cxx: Func_void_std__string_bool {
   override fun invoke(token: String, done: Boolean): Unit
     = invoke_cxx(token,done)
 
-  @FastNative
   private external fun invoke_cxx(token: String, done: Boolean): Unit
 }
 

package/nitrogen/generated/android/kotlin/com/margelo/nitro/dev/litert/litertlm/HybridLiteRTLMSpec.kt

@@ -10,6 +10,7 @@ package com.margelo.nitro.dev.litert.litertlm
 import androidx.annotation.Keep
 import com.facebook.jni.HybridData
 import com.facebook.proguard.annotations.DoNotStrip
+import dalvik.annotation.optimization.FastNative
 import com.margelo.nitro.core.Promise
 import com.margelo.nitro.core.HybridObject
 
@@ -157,6 +158,7 @@ abstract class HybridLiteRTLMSpec: HybridObject() {
   @Keep
   protected open class CxxPart(javaPart: HybridLiteRTLMSpec): HybridObject.CxxPart(javaPart) {
     // C++ JHybridLiteRTLMSpec::CxxPart::initHybrid(...)
+    @FastNative
     external override fun initHybrid(): HybridData
   }
   override fun createCxxPart(): CxxPart {

package/nitrogen/generated/android/kotlin/com/margelo/nitro/dev/litert/litertlm/HybridModelStoreSpec.kt

@@ -10,6 +10,7 @@ package com.margelo.nitro.dev.litert.litertlm
 import androidx.annotation.Keep
 import com.facebook.jni.HybridData
 import com.facebook.proguard.annotations.DoNotStrip
+import dalvik.annotation.optimization.FastNative
 import com.margelo.nitro.core.Promise
 import com.margelo.nitro.core.HybridObject
 
@@ -64,6 +65,7 @@ abstract class HybridModelStoreSpec: HybridObject() {
   @Keep
   protected open class CxxPart(javaPart: HybridModelStoreSpec): HybridObject.CxxPart(javaPart) {
     // C++ JHybridModelStoreSpec::CxxPart::initHybrid(...)
+    @FastNative
     external override fun initHybrid(): HybridData
   }
   override fun createCxxPart(): CxxPart {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dr33m/react-native-litert-lm",
-  "version": "0.5.5",
+  "version": "0.5.6",
   "litertLm": {
     "version": "0.12.0",
     "androidMavenVersion": "0.12.0",