llama-cpp-pro 0.2.2

package/build-native.sh

@@ -0,0 +1,280 @@
+#!/bin/bash
+
+# Build script for llama-cpp Capacitor plugin
+# This script compiles the native llama.cpp library for iOS and Android
+
+set -e
+
+# Colors for output
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+NC='\033[0m' # No Color
+
+# Function to print colored output
+print_status() {
+    echo -e "${BLUE}[INFO]${NC} $1"
+}
+
+print_success() {
+    echo -e "${GREEN}[SUCCESS]${NC} $1"
+}
+
+print_warning() {
+    echo -e "${YELLOW}[WARNING]${NC} $1"
+}
+
+print_error() {
+    echo -e "${RED}[ERROR]${NC} $1"
+}
+
+bytes_to_mb() {
+    awk -v bytes="$1" 'BEGIN { printf "%.2f", bytes / 1048576 }'
+}
+
+# Check if we're on macOS for iOS builds
+check_macos() {
+    if [[ "$OSTYPE" != "darwin"* ]]; then
+        print_warning "iOS builds require macOS. Skipping iOS build."
+        return 1
+    fi
+    return 0
+}
+
+# Check if Android SDK is available
+check_android_sdk() {
+    if [ -z "$ANDROID_HOME" ] && [ -z "$ANDROID_SDK_ROOT" ]; then
+        print_warning "Android SDK not found. Please set ANDROID_HOME or ANDROID_SDK_ROOT."
+        print_warning "Skipping Android build."
+        return 1
+    fi
+    return 0
+}
+
+# Detect NDK path: ndk/<version>/build/cmake/android.toolchain.cmake
+# Uses latest versioned NDK under $ANDROID_SDK/ndk/ (e.g. ndk/29.0.13113456).
+detect_ndk() {
+    local sdk="${ANDROID_HOME:-$ANDROID_SDK_ROOT}"
+    local ndk_root="$sdk/ndk"
+    if [ ! -d "$ndk_root" ]; then
+        echo ""
+        return 1
+    fi
+    local latest=""
+    local latest_ver=0
+    for v in "$ndk_root"/*; do
+        [ -d "$v" ] || continue
+        local base=$(basename "$v")
+        if [[ "$base" =~ ^[0-9] ]]; then
+            local ver=$(echo "$base" | sed 's/[^0-9]//g' | head -c 10)
+            ver=${ver:-0}
+            if [ "$ver" -gt "$latest_ver" ] 2>/dev/null; then
+                latest_ver=$ver
+                latest=$v
+            fi
+        fi
+    done
+    if [ -n "$latest" ] && [ -f "$latest/build/cmake/android.toolchain.cmake" ]; then
+        echo "$latest"
+        return 0
+    fi
+    echo ""
+    return 1
+}
+
+# Build iOS library
+build_ios() {
+    print_status "Building iOS library..."
+    
+    if ! check_macos; then
+        return 1
+    fi
+    
+    # Always start from a clean iOS build directory to avoid stale Xcode settings
+    rm -rf ios/build
+    mkdir -p ios/build
+    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 (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: $(pwd)/llama-cpp.framework"
+        
+        # Binary location: CMake may produce flat (llama-cpp) or Versions/A/ layout
+        BINARY=""
+        if [ -f "llama-cpp.framework/llama-cpp" ]; then
+            BINARY="llama-cpp.framework/llama-cpp"
+        elif [ -f "llama-cpp.framework/Versions/A/llama-cpp" ]; then
+            BINARY="llama-cpp.framework/Versions/A/llama-cpp"
+        fi
+        if [ -n "$BINARY" ]; then
+            if xcrun strip -x -S "$BINARY" 2>/dev/null; then
+                print_status "Stripped debug symbols from iOS framework"
+            fi
+        fi
+        
+        # Build flat framework (single binary, no duplication) for npm publishing
+        if [ -z "$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 ../..
+        return 1
+    fi
+    
+    cd ../..
+}
+
+# Build Android library
+build_android() {
+    print_status "Building Android library..."
+    
+    if ! check_android_sdk; then
+        return 1
+    fi
+    
+    ANDROID_NDK=$(detect_ndk)
+    if [ -z "$ANDROID_NDK" ]; then
+        print_error "Android NDK not found. Install NDK via Android Studio (SDK Manager → NDK)."
+        print_error "Expected: \$ANDROID_HOME/ndk/<version>/build/cmake/android.toolchain.cmake"
+        return 1
+    fi
+    print_status "Using NDK: $ANDROID_NDK"
+    
+    TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake"
+    if [ ! -f "$TOOLCHAIN_FILE" ]; then
+        print_error "Toolchain file not found: $TOOLCHAIN_FILE"
+        return 1
+    fi
+    
+    rm -rf android/build
+    mkdir -p android/build
+    cd android/build
+    
+    # Build only arm64-v8a to minimize app store size.
+    # build.gradle uses abiFilters 'arm64-v8a'; armeabi-v7a is not included in the app.
+    # Skipping armeabi-v7a saves ~25–48 MB in package and keeps app size minimal.
+    arch="arm64-v8a"
+    print_status "Building for $arch (only ABI shipped in app)..."
+    rm -rf CMakeCache.txt CMakeFiles Makefile cmake_install.cmake 2>/dev/null || true
+    find . -maxdepth 1 -name '*.so' -delete 2>/dev/null || true
+    
+    cmake ../src/main \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DANDROID_ABI=$arch \
+        -DANDROID_PLATFORM=android-21 \
+        -DCMAKE_TOOLCHAIN_FILE="$TOOLCHAIN_FILE" \
+        -DANDROID_STL=c++_shared
+    
+    JOBS=$(sysctl -n hw.ncpu 2>/dev/null || nproc 2>/dev/null || echo 4)
+    cmake --build . --config Release -- -j"$JOBS"
+    
+    FINAL_SO="../src/main/jniLibs/$arch/libllama-cpp-arm64.so"
+    if [ -f "$FINAL_SO" ]; then
+        PREBUILT="$ANDROID_NDK/toolchains/llvm/prebuilt"
+        TOOL_BIN=""
+        if [ -d "$PREBUILT/darwin-x86_64" ]; then
+            TOOL_BIN="$PREBUILT/darwin-x86_64/bin"
+        elif [ -d "$PREBUILT/darwin-aarch64" ]; then
+            TOOL_BIN="$PREBUILT/darwin-aarch64/bin"
+        fi
+        if [ -n "$TOOL_BIN" ]; then
+            STRIP_TOOL="$TOOL_BIN/llvm-strip"
+            OBJCOPY_TOOL="$TOOL_BIN/llvm-objcopy"
+        else
+            STRIP_TOOL=$(find "$ANDROID_NDK/toolchains" -name "llvm-strip" -type f 2>/dev/null | head -1)
+            OBJCOPY_TOOL=$(find "$ANDROID_NDK/toolchains" -name "llvm-objcopy" -type f 2>/dev/null | head -1)
+        fi
+
+        BEFORE_SIZE=$(wc -c < "$FINAL_SO")
+        SYMBOLS_DIR="../symbols/$arch"
+        SYMBOL_FILE="$SYMBOLS_DIR/libllama-cpp-arm64.so.debug"
+        mkdir -p "$SYMBOLS_DIR"
+
+        if [ -f "$OBJCOPY_TOOL" ]; then
+            "$OBJCOPY_TOOL" --only-keep-debug "$FINAL_SO" "$SYMBOL_FILE"
+            print_status "Saved Android debug symbols to: $SYMBOL_FILE"
+        else
+            cp "$FINAL_SO" "$SYMBOL_FILE"
+            print_warning "llvm-objcopy not found; copied unstripped library as debug symbol file"
+        fi
+
+        if [ -f "$STRIP_TOOL" ]; then
+            "$STRIP_TOOL" --strip-unneeded "$FINAL_SO"
+            if [ -f "$OBJCOPY_TOOL" ]; then
+                "$OBJCOPY_TOOL" --add-gnu-debuglink="$SYMBOL_FILE" "$FINAL_SO" || true
+            fi
+            AFTER_SIZE=$(wc -c < "$FINAL_SO")
+            print_status "Stripped Android release library (--strip-unneeded): $(bytes_to_mb "$BEFORE_SIZE") MB -> $(bytes_to_mb "$AFTER_SIZE") MB"
+        fi
+    else
+        print_warning "Expected Android library not found at $FINAL_SO"
+    fi
+    print_success "Built for $arch"
+    
+    print_success "Android library built successfully"
+    cd ../..
+}
+
+# Main build function
+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"
+        exit 1
+    fi
+    
+    if ! command -v make &> /dev/null; then
+        print_error "Make is required but not installed"
+        exit 1
+    fi
+    
+    # Build iOS
+    if check_macos; then
+        build_ios
+    fi
+    
+    # Build Android
+    if check_android_sdk; then
+        build_android
+    fi
+    
+    print_success "Build completed successfully!"
+}
+
+# Run main function (ignore any args npm may pass)
+main

package/cmake/desktop-metal-embed.cmake

@@ -0,0 +1,30 @@
+# Embed ggml-metal.metal for desktop sidecar builds.
+# Same as ios/metal-embed.cmake but paths are relative to this repo root.
+
+set(SOURCE_DIR "${ANNADATA_CPP_DIR}")
+set(METALLIB_COMMON "${SOURCE_DIR}/ggml-common.h")
+set(METALLIB_SOURCE "${SOURCE_DIR}/ggml-metal.metal")
+set(METALLIB_IMPL "${SOURCE_DIR}/ggml-metal-impl.h")
+set(METALLIB_EMBED_SCRIPT "${CMAKE_CURRENT_LIST_DIR}/../ios/embed-metal-shaders.sh")
+
+file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
+
+set(METALLIB_EMBED_ASM "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+set(METALLIB_SOURCE_EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
+
+add_custom_command(
+    OUTPUT "${METALLIB_EMBED_ASM}"
+    COMMAND "${METALLIB_EMBED_SCRIPT}"
+        "${METALLIB_COMMON}"
+        "${METALLIB_SOURCE}"
+        "${METALLIB_IMPL}"
+        "${METALLIB_EMBED_ASM}"
+        "${METALLIB_SOURCE_EMBED}"
+    DEPENDS
+        "${METALLIB_EMBED_SCRIPT}"
+        "${METALLIB_COMMON}"
+        "${METALLIB_SOURCE}"
+        "${METALLIB_IMPL}"
+    COMMENT "Generate assembly for embedded Metal shader source (desktop sidecar)"
+    VERBATIM
+)

package/cmake/desktop-sources.cmake

@@ -0,0 +1,100 @@
+# Shared llama.cpp / cap-* source list for desktop sidecar builds.
+# Included by sidecar/CMakeLists.txt.
+
+set(ANNADATA_CPP_DIR "${CMAKE_CURRENT_LIST_DIR}/../cpp")
+
+set(ANNADATA_DESKTOP_CORE_SOURCES
+    ${ANNADATA_CPP_DIR}/ggml.c
+    ${ANNADATA_CPP_DIR}/ggml-alloc.c
+    ${ANNADATA_CPP_DIR}/ggml-backend.cpp
+    ${ANNADATA_CPP_DIR}/ggml-backend-reg.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/amx/amx.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/amx/mmq.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/ggml-cpu.c
+    ${ANNADATA_CPP_DIR}/ggml-cpu/ggml-cpu.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/quants.c
+    ${ANNADATA_CPP_DIR}/ggml-cpu/traits.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/repack.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/unary-ops.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/binary-ops.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/vec.cpp
+    ${ANNADATA_CPP_DIR}/ggml-cpu/ops.cpp
+    ${ANNADATA_CPP_DIR}/ggml-opt.cpp
+    ${ANNADATA_CPP_DIR}/ggml-threading.cpp
+    ${ANNADATA_CPP_DIR}/ggml-quants.c
+    ${ANNADATA_CPP_DIR}/gguf.cpp
+    ${ANNADATA_CPP_DIR}/log.cpp
+    ${ANNADATA_CPP_DIR}/llama-impl.cpp
+    ${ANNADATA_CPP_DIR}/chat-parser.cpp
+    ${ANNADATA_CPP_DIR}/json-partial.cpp
+    ${ANNADATA_CPP_DIR}/regex-partial.cpp
+    ${ANNADATA_CPP_DIR}/tools/mtmd/mtmd.cpp
+    ${ANNADATA_CPP_DIR}/tools/mtmd/mtmd-audio.cpp
+    ${ANNADATA_CPP_DIR}/tools/mtmd/clip.cpp
+    ${ANNADATA_CPP_DIR}/tools/mtmd/mtmd-helper.cpp
+    ${ANNADATA_CPP_DIR}/llama-grammar.cpp
+    ${ANNADATA_CPP_DIR}/llama-sampling.cpp
+    ${ANNADATA_CPP_DIR}/llama-vocab.cpp
+    ${ANNADATA_CPP_DIR}/llama-adapter.cpp
+    ${ANNADATA_CPP_DIR}/llama-chat.cpp
+    ${ANNADATA_CPP_DIR}/llama-context.cpp
+    ${ANNADATA_CPP_DIR}/llama-arch.cpp
+    ${ANNADATA_CPP_DIR}/llama-batch.cpp
+    ${ANNADATA_CPP_DIR}/llama-cparams.cpp
+    ${ANNADATA_CPP_DIR}/llama-hparams.cpp
+    ${ANNADATA_CPP_DIR}/llama.cpp
+    ${ANNADATA_CPP_DIR}/llama-model.cpp
+    ${ANNADATA_CPP_DIR}/llama-model-loader.cpp
+    ${ANNADATA_CPP_DIR}/llama-model-saver.cpp
+    ${ANNADATA_CPP_DIR}/llama-kv-cache.cpp
+    ${ANNADATA_CPP_DIR}/llama-kv-cache-iswa.cpp
+    ${ANNADATA_CPP_DIR}/llama-memory-hybrid.cpp
+    ${ANNADATA_CPP_DIR}/llama-memory-recurrent.cpp
+    ${ANNADATA_CPP_DIR}/llama-mmap.cpp
+    ${ANNADATA_CPP_DIR}/llama-memory.cpp
+    ${ANNADATA_CPP_DIR}/llama-io.cpp
+    ${ANNADATA_CPP_DIR}/llama-graph.cpp
+    ${ANNADATA_CPP_DIR}/sampling.cpp
+    ${ANNADATA_CPP_DIR}/unicode-data.cpp
+    ${ANNADATA_CPP_DIR}/unicode.cpp
+    ${ANNADATA_CPP_DIR}/common.cpp
+    ${ANNADATA_CPP_DIR}/chat.cpp
+    ${ANNADATA_CPP_DIR}/json-schema-to-grammar.cpp
+    ${ANNADATA_CPP_DIR}/anyascii.c
+    ${ANNADATA_CPP_DIR}/cap-llama.cpp
+    ${ANNADATA_CPP_DIR}/cap-completion.cpp
+    ${ANNADATA_CPP_DIR}/cap-tts.cpp
+    ${ANNADATA_CPP_DIR}/cap-embedding.cpp
+    ${ANNADATA_CPP_DIR}/cap-ios-bridge.cpp
+    ${ANNADATA_CPP_DIR}/cap-native-server.cpp
+    ${ANNADATA_CPP_DIR}/vendor/cpp-httplib/httplib.cpp
+)
+
+# Architecture-specific CPU kernels
+set(_DESKTOP_IS_ARM64 OFF)
+if(APPLE)
+    if(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
+        set(_DESKTOP_IS_ARM64 ON)
+    elseif(NOT CMAKE_OSX_ARCHITECTURES OR CMAKE_OSX_ARCHITECTURES STREQUAL "")
+        execute_process(COMMAND uname -m OUTPUT_VARIABLE _UNAME_M OUTPUT_STRIP_TRAILING_WHITESPACE)
+        if(_UNAME_M STREQUAL "arm64")
+            set(_DESKTOP_IS_ARM64 ON)
+        endif()
+    endif()
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
+    set(_DESKTOP_IS_ARM64 ON)
+endif()
+
+if(_DESKTOP_IS_ARM64)
+    list(APPEND ANNADATA_DESKTOP_CORE_SOURCES
+        ${ANNADATA_CPP_DIR}/ggml-cpu/arch/arm/quants.c
+        ${ANNADATA_CPP_DIR}/ggml-cpu/arch/arm/repack.cpp
+        ${ANNADATA_CPP_DIR}/ggml-cpu/arch/arm/cpu-feats.cpp
+    )
+else()
+    list(APPEND ANNADATA_DESKTOP_CORE_SOURCES
+        ${ANNADATA_CPP_DIR}/ggml-cpu/arch/x86/quants.c
+        ${ANNADATA_CPP_DIR}/ggml-cpu/arch/x86/repack.cpp
+        ${ANNADATA_CPP_DIR}/ggml-cpu/arch/x86/cpu-feats.cpp
+    )
+endif()

package/cmake/ggml-backends.cmake

@@ -0,0 +1,44 @@
+# Optional ggml GPU backend plugin builds
+#
+# Requires LLAMA_CPP_UPSTREAM pointing at a full upstream llama.cpp checkout.
+# Produces libggml-vulkan.so, libggml-cuda.so, libggml-hip.so next to the sidecar.
+#
+# Usage:
+#   cmake -B sidecar/build -DLLAMA_CPP_UPSTREAM=/path/to/llama.cpp -DSIDECAR_VARIANT=vulkan-openblas
+#   cmake --build sidecar/build --target ggml-backends
+
+if(NOT LLAMA_CPP_UPSTREAM OR NOT EXISTS "${LLAMA_CPP_UPSTREAM}/ggml/CMakeLists.txt")
+    message(STATUS "ggml-backends: LLAMA_CPP_UPSTREAM not set — skipping GPU plugin build")
+    return()
+endif()
+
+message(STATUS "ggml-backends: building plugins from ${LLAMA_CPP_UPSTREAM}")
+
+set(GGML_BUILD_TESTS OFF CACHE BOOL "" FORCE)
+set(GGML_BUILD_EXAMPLES OFF CACHE BOOL "" FORCE)
+set(BUILD_SHARED_LIBS ON CACHE BOOL "" FORCE)
+
+set(_GGML_VULKAN OFF)
+set(_GGML_CUDA OFF)
+set(_GGML_HIP OFF)
+
+if(SIDECAR_VARIANT MATCHES "vulkan")
+    set(_GGML_VULKAN ON)
+elseif(SIDECAR_VARIANT STREQUAL "cuda")
+    set(_GGML_CUDA ON)
+elseif(SIDECAR_VARIANT STREQUAL "rocm")
+    set(_GGML_HIP ON)
+elseif(SIDECAR_VARIANT STREQUAL "vulkan-openblas")
+    set(_GGML_VULKAN ON)
+endif()
+
+set(GGML_VULKAN ${_GGML_VULKAN} CACHE BOOL "" FORCE)
+set(GGML_CUDA ${_GGML_CUDA} CACHE BOOL "" FORCE)
+set(GGML_HIP ${_GGML_HIP} CACHE BOOL "" FORCE)
+
+add_subdirectory("${LLAMA_CPP_UPSTREAM}/ggml" "${CMAKE_BINARY_DIR}/ggml-upstream")
+
+add_custom_target(ggml-backends ALL
+    DEPENDS ggml
+    COMMENT "ggml GPU backend plugins built into ${CMAKE_BINARY_DIR}/bin"
+)

package/cpp/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023-2024 The ggml authors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/cpp/README.md

@@ -0,0 +1,15 @@
+# Note
+
+- The **project-specific** files in this folder (never overwritten by bootstrap) are:
+  - `cap-llama.cpp` / `cap-llama.h` – Capacitor bridge (context, model load)
+  - `cap-completion.cpp` / `cap-completion.h` – completion
+  - `cap-tts.cpp` / `cap-tts.h` – TTS
+  - `cap-embedding.cpp` / `cap-embedding.h` – embeddings
+  - `cap-mtmd.hpp` – multimodal/vision
+  - `tools/mtmd/` – multimodal/vision tooling
+- All other sources are synced from [llama.cpp](https://github.com/ggerganov/llama.cpp).
+- **Update native source** (e.g. for a newer llama.cpp or vision support): run from repo root:
+  ```bash
+  ./scripts/bootstrap.sh [branch-or-tag-or-commit]
+  ```
+  See [scripts/bootstrap.sh](../scripts/bootstrap.sh) and [docs/IOS_IMPLEMENTATION_GUIDE.md](../docs/IOS_IMPLEMENTATION_GUIDE.md).