llama-cpp-capacitor 0.1.3 → 0.1.4

package/LlamaCppCapacitor.podspec

@@ -17,4 +17,7 @@ Pod::Spec.new do |s|
   
   # Include the native llama-cpp framework
   s.vendored_frameworks = 'ios/Frameworks/llama-cpp.framework'
+  s.pod_target_xcconfig = {
+    'FRAMEWORK_SEARCH_PATHS' => '$(inherited) "$(PODS_TARGET_SRCROOT)/ios/Frameworks"'
+  }
 end

package/build-native.sh

@@ -93,36 +93,53 @@ build_ios() {
     cd ios/build
     
     # Configure with CMake
+    # IMPORTANT: CMAKE_OSX_SYSROOT=iphoneos ensures we build for iOS, not macOS.
+    # Without it, the framework would be built for macOS and fail to link in an iOS app.
     # IMPORTANT: build iOS framework as **ARM64-only**.
     # Including x86_64 here makes CMake/Xcode try to link an x86_64 slice,
     # but we only compile ARM-specific kernels (arch/arm), which leads to
     # undefined symbols like lm_ggml_gemm_* for x86_64.
     cmake .. \
         -DCMAKE_BUILD_TYPE=Release \
+        -DCMAKE_OSX_SYSROOT=iphoneos \
         -DCMAKE_OSX_ARCHITECTURES="arm64" \
         -DCMAKE_OSX_DEPLOYMENT_TARGET=13.0 \
         -DCMAKE_XCODE_ATTRIBUTE_ENABLE_BITCODE=NO
     
-    # Build
-    cmake --build . --config Release
+    # Build (parallel: -j uses available CPU cores)
+    JOBS=$(sysctl -n hw.ncpu 2>/dev/null || nproc 2>/dev/null || echo 4)
+    cmake --build . --config Release -- -j"$JOBS"
     
     # CMake builds the framework directly (FRAMEWORK TRUE in CMakeLists.txt)
     # Verify the framework was created
     if [ -d "llama-cpp.framework" ]; then
         print_success "iOS framework built successfully at: $(pwd)/llama-cpp.framework"
         
-        # Strip debug symbols to reduce app store size (~0.5–1 MB)
-        BINARY="llama-cpp.framework/Versions/A/llama-cpp"
-        if [ -f "$BINARY" ]; then
+        # Binary location: CMake may produce flat (llama-cpp) or Versions/A/ layout
+        BINARY=""
+        if [ -f "llama-cpp.framework/llama-cpp" ]; then
+            BINARY="llama-cpp.framework/llama-cpp"
+        elif [ -f "llama-cpp.framework/Versions/A/llama-cpp" ]; then
+            BINARY="llama-cpp.framework/Versions/A/llama-cpp"
+        fi
+        if [ -n "$BINARY" ]; then
             if xcrun strip -x -S "$BINARY" 2>/dev/null; then
                 print_status "Stripped debug symbols from iOS framework"
             fi
         fi
         
-        # Copy framework to package location for npm publishing
-        mkdir -p ../Frameworks
-        cp -R llama-cpp.framework ../Frameworks/
-        print_success "iOS framework copied to ios/Frameworks/ for npm package"
+        # Build flat framework (single binary, no duplication) for npm publishing
+        if [ -z "$BINARY" ] || [ ! -f "$BINARY" ]; then
+            print_error "iOS framework binary not found"
+            cd ../..
+            return 1
+        fi
+        rm -rf ../Frameworks/llama-cpp.framework
+        mkdir -p ../Frameworks/llama-cpp.framework/Resources
+        cp "$BINARY" ../Frameworks/llama-cpp.framework/llama-cpp
+        [ -f llama-cpp.framework/Info.plist ] && cp llama-cpp.framework/Info.plist ../Frameworks/llama-cpp.framework/
+        [ -f llama-cpp.framework/Versions/A/Resources/Info.plist ] && cp llama-cpp.framework/Versions/A/Resources/Info.plist ../Frameworks/llama-cpp.framework/Resources/
+        print_success "iOS framework copied to ios/Frameworks/ for npm package (flat, no duplication)"
     else
         print_error "iOS framework not found after build"
         cd ../..
@@ -173,7 +190,8 @@ build_android() {
         -DCMAKE_TOOLCHAIN_FILE="$TOOLCHAIN_FILE" \
         -DANDROID_STL=c++_shared
     
-    cmake --build . --config Release
+    JOBS=$(sysctl -n hw.ncpu 2>/dev/null || nproc 2>/dev/null || echo 4)
+    cmake --build . --config Release -- -j"$JOBS"
     
     mkdir -p ../src/main/jniLibs/$arch
     if [ -f "libllama-cpp-arm64.so" ]; then
@@ -201,6 +219,10 @@ build_android() {
 main() {
     print_status "Starting llama-cpp Capacitor plugin build..."
     
+    # Always start clean - remove previous build outputs
+    rm -rf ios/build ios/Frameworks android/build
+    print_status "Cleaned build directories"
+    
     # Check dependencies
     if ! command -v cmake &> /dev/null; then
         print_error "CMake is required but not installed"
@@ -225,5 +247,5 @@ main() {
     print_success "Build completed successfully!"
 }
 
-# Run main function
-main "$@"
+# Run main function (ignore any args npm may pass)
+main

package/ios/Sources/LlamaCppPlugin/LlamaCpp.swift

@@ -150,6 +150,8 @@ struct MinjaCaps {
 // MARK: - Main Implementation
 @objc public class LlamaCpp: NSObject {
     private var contexts: [Int: LlamaContext] = [:]
+    private var nativeContexts: [Int64: UnsafeMutableRawPointer] = [:]
+    private var contextIdToNative: [Int: Int64] = [:]
     private var contextCounter: Int = 0
     private var contextLimit: Int = 10
     private var nativeLogEnabled: Bool = false
@@ -262,18 +264,16 @@ struct MinjaCaps {
         
         let nativeContextId = initFunc(modelPath, paramsJson.cString(using: .utf8)!)
         if nativeContextId > 0 {
-            // Store the native context pointer (the ID is used as the pointer value)
-            // Note: In a real implementation, the native function would return the actual pointer
-            // For now, we use the contextId as the pointer identifier
-            contexts[Int64(contextId)] = UnsafeMutableRawPointer(bitPattern: Int(nativeContextId))
+            // Store the LlamaContext for Swift bookkeeping
+            contexts[contextId] = context
+            // Store the native context pointer and mapping for C layer
+            let nativePtr = UnsafeMutableRawPointer(bitPattern: Int(nativeContextId))
+            nativeContexts[nativeContextId] = nativePtr
+            contextIdToNative[contextId] = nativeContextId
             
             // Register with embedding system if available
-            // The C layer needs the actual llama_cap_context pointer, which should come from the native init
-            // For now, we'll register with the contextId - the C layer will need to look it up
-            if let registerFunc = registerEmbeddingContextFunc {
-                // Note: The actual context pointer should come from the native initContext function
-                // This is a placeholder - the real implementation needs to get the actual pointer
-                registerFunc(Int64(contextId), contexts[Int64(contextId)]!)
+            if let registerFunc = registerEmbeddingContextFunc, let ptr = nativePtr {
+                registerFunc(nativeContextId, ptr)
             }
         } else {
             completion(.failure(.operationFailed("Failed to initialize native context")))
@@ -322,34 +322,40 @@ struct MinjaCaps {
     }
     
     func releaseContext(contextId: Int, completion: @escaping (LlamaResult<Void>) -> Void) {
-        guard let contextPtr = contexts[contextId] else {
+        guard contexts[contextId] != nil else {
             completion(.failure(.contextNotFound))
             return
         }
         
+        let nativeId = contextIdToNative[contextId] ?? Int64(contextId)
+        
         // Unregister from embedding system if available
         if let unregisterFunc = unregisterEmbeddingContextFunc {
-            unregisterFunc(Int64(contextId))
+            unregisterFunc(nativeId)
         }
         
         // Call native release function
         if let releaseFunc = releaseContextFunc {
-            releaseFunc(Int64(contextId))
+            releaseFunc(nativeId)
         }
         
         contexts.removeValue(forKey: contextId)
+        nativeContexts.removeValue(forKey: nativeId)
+        contextIdToNative.removeValue(forKey: contextId)
         completion(.success(()))
     }
     
     func releaseAllContexts(completion: @escaping (LlamaResult<Void>) -> Void) {
         // Unregister all contexts from embedding system
         if let unregisterFunc = unregisterEmbeddingContextFunc {
-            for contextId in contexts.keys {
-                unregisterFunc(Int64(contextId))
+            for (_, nativeId) in contextIdToNative {
+                unregisterFunc(nativeId)
             }
         }
         
         contexts.removeAll()
+        nativeContexts.removeAll()
+        contextIdToNative.removeAll()
         completion(.success(()))
     }
     

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "llama-cpp-capacitor",
-  "version": "0.1.3",
+  "version": "0.1.4",
   "description": "A native Capacitor plugin that embeds llama.cpp directly into mobile apps, enabling offline AI inference with chat-first API design. Complete iOS and Android support: text generation, chat, multimodal, TTS, LoRA, embeddings, and more.",
   "main": "dist/plugin.cjs.js",
   "type": "module",
@@ -72,6 +72,7 @@
     "build:all": "npm run build && npm run build:native",
     "build:ios": "cd ios && cmake -B build -S . && cmake --build build --config Release",
     "build:android": "cd android && gradlew.bat assembleRelease",
+    "prepack": "npm run build:native",
     "pack": "npm run build && npm pack --dry-run",
     "pack:full": "npm run build:all && npm pack --dry-run",
     "test": "jest",
@@ -84,7 +85,7 @@
     "clean:native": "rimraf ios/build ios/Frameworks android/build android/src/main/jniLibs",
     "clean:test": "rimraf test/output test/coverage",
     "watch": "tsc --watch",
-    "prepublishOnly": "npm run build"
+    "prepublishOnly": "npm run build && npm run build:native"
   },
   "devDependencies": {
     "@capacitor/android": "^7.0.0",