@runanywhere/llamacpp 0.18.0 → 0.19.13

package/android/CMakeLists.txt

@@ -1,5 +1,24 @@
+# =============================================================================
+# COUPLING NOTE: This file references librac_commons.so + RAC headers from the
+# sibling @runanywhere/core npm package via a hard-coded relative path:
+#   ../../core/android/src/main/jniLibs/${ANDROID_ABI}/librac_commons.so
+#   ../../core/android/src/main/jniLibs/include/
+# (computed via RN_NODE_MODULES → @runanywhere/core/android/src/main/jniLibs/).
+#
+# This is a known fragility — if the package directory structure changes,
+# update the path here. Future improvement: migrate to Android Prefab module
+# consumption (build.gradle: `prefab { commons { ... } }`) so the dependency
+# is declared at the Gradle layer rather than the CMake layer.
+#
+# Prefab adoption was attempted but deferred because librac_commons.so is a
+# pre-built binary downloaded by core's Gradle task (not produced by core's
+# externalNativeBuild), so AGP's `prefabPublishing` publisher cannot
+# auto-export it without additional manifest wiring. The relative-path
+# coupling is therefore documented and intentionally preserved.
+# =============================================================================
 project(runanywherellama)
-cmake_minimum_required(VERSION 3.9.0)
+# Must match sdk/runanywhere-commons/VERSIONS::MIN_CMAKE_VERSION (3.24).
+cmake_minimum_required(VERSION 3.24)
 
 set(PACKAGE_NAME runanywherellama)
 set(CMAKE_VERBOSE_MAKEFILE ON)
@@ -9,7 +28,7 @@ set(CMAKE_CXX_STANDARD 20)
 # 16KB Page Alignment for Android 15+ (API 35) Compliance
 # Required starting November 1, 2025 for Google Play submissions
 # =============================================================================
-set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -Wl,-z,max-page-size=16384")
+set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -Wl,-z,max-page-size=16384 -Wl,-z,common-page-size=16384")
 
 # Path to pre-built native libraries (downloaded from runanywhere-binaries)
 set(JNILIB_DIR ${CMAKE_SOURCE_DIR}/src/main/jniLibs/${ANDROID_ABI})
@@ -19,12 +38,8 @@ set(JNILIB_DIR ${CMAKE_SOURCE_DIR}/src/main/jniLibs/${ANDROID_ABI})
 # Downloaded via Gradle downloadNativeLibs task
 # =============================================================================
 if(NOT EXISTS "${JNILIB_DIR}/librac_backend_llamacpp.so")
-    message(WARNING "[RunAnywhereLlama] RABackendLlamaCPP not found for ${ANDROID_ABI} at ${JNILIB_DIR}/librac_backend_llamacpp.so\n"
-                    "This ABI will not be functional. To fix, run: ./gradlew :runanywhere_llamacpp:downloadNativeLibs\n"
-                    "Or set reactNativeArchitectures=arm64-v8a in gradle.properties to skip this ABI.")
-    file(WRITE "${CMAKE_CURRENT_BINARY_DIR}/stub.cpp" "// Stub for missing ABI ${ANDROID_ABI}")
-    add_library(${PACKAGE_NAME} SHARED "${CMAKE_CURRENT_BINARY_DIR}/stub.cpp")
-    return()
+    message(FATAL_ERROR "[RunAnywhereLlama] RABackendLlamaCPP not found for ${ANDROID_ABI} at ${JNILIB_DIR}/librac_backend_llamacpp.so\n"
+                        "Run ./scripts/build/build-core-android.sh from the repo root or set reactNativeArchitectures to an ABI with staged natives.")
 endif()
 
 add_library(rac_backend_llamacpp SHARED IMPORTED)
@@ -34,15 +49,9 @@ set_target_properties(rac_backend_llamacpp PROPERTIES
 )
 message(STATUS "[RunAnywhereLlama] Found RABackendLlamaCPP at ${JNILIB_DIR}/librac_backend_llamacpp.so")
 
-# =============================================================================
-# Source files - Llama bridges
-# =============================================================================
-file(GLOB BRIDGE_SOURCES "../cpp/bridges/*.cpp")
-
 add_library(${PACKAGE_NAME} SHARED
     src/main/cpp/cpp-adapter.cpp
     ../cpp/HybridRunAnywhereLlama.cpp
-    ${BRIDGE_SOURCES}
 )
 
 # =============================================================================
@@ -70,13 +79,19 @@ include(${CMAKE_SOURCE_DIR}/../nitrogen/generated/android/runanywherellama+autol
 # =============================================================================
 # Get core package include dir (for rac/*.h headers)
 get_filename_component(RN_NODE_MODULES "${CMAKE_SOURCE_DIR}/../../.." ABSOLUTE)
-set(CORE_INCLUDE_DIR "${RN_NODE_MODULES}/@runanywhere/core/android/src/main/include")
+set(CORE_INCLUDE_DIR "${RN_NODE_MODULES}/@runanywhere/core/android/src/main/jniLibs/include")
 set(CORE_JNILIB_DIR "${RN_NODE_MODULES}/@runanywhere/core/android/src/main/jniLibs/${ANDROID_ABI}")
 
 include_directories(
     "src/main/cpp"
+    # Package-local hybrid headers (HybridRunAnywhereLlama.hpp etc.) — needed by the
+    # nitrogen-generated OnLoad + cpp-adapter. The forked rac_llm_llamacpp.h that used
+    # to live here was deleted (canonical commons header resolves via CORE_INCLUDE_DIR).
     "../cpp"
-    "../cpp/bridges"
+    # nlohmann/json (header-only, vendored by sibling @runanywhere/core package).
+    # Sharing avoids re-fetching the dependency and keeps both packages on the
+    # same JSON version.
+    "${CMAKE_SOURCE_DIR}/../../core/cpp/third_party"
     "${CMAKE_SOURCE_DIR}/include"
     # RAC API headers from core package (flat access for all subdirectories)
     "${CORE_INCLUDE_DIR}"
@@ -88,8 +103,10 @@ include_directories(
     "${CORE_INCLUDE_DIR}/rac/features/stt"
     "${CORE_INCLUDE_DIR}/rac/features/tts"
     "${CORE_INCLUDE_DIR}/rac/features/vad"
+    "${CORE_INCLUDE_DIR}/rac/features/vlm"
     "${CORE_INCLUDE_DIR}/rac/features/voice_agent"
     "${CORE_INCLUDE_DIR}/rac/features/platform"
+    "${CORE_INCLUDE_DIR}/rac/backends"
     "${CORE_INCLUDE_DIR}/rac/infrastructure"
     "${CORE_INCLUDE_DIR}/rac/infrastructure/device"
     "${CORE_INCLUDE_DIR}/rac/infrastructure/download"
@@ -124,12 +141,12 @@ target_link_libraries(
     ${PACKAGE_NAME}
     ${LOG_LIB}
     android
+    c++_shared
     rac_commons
     rac_backend_llamacpp
 )
 
-# HAS_LLAMACPP and HAS_RACOMMONS are always defined since backends are required
-target_compile_definitions(${PACKAGE_NAME} PRIVATE HAS_LLAMACPP=1 HAS_RACOMMONS=1)
-
 # 16KB page alignment - MUST be on target for Android 15+ compliance
-target_link_options(${PACKAGE_NAME} PRIVATE -Wl,-z,max-page-size=16384)
+target_link_options(${PACKAGE_NAME} PRIVATE
+    -Wl,-z,max-page-size=16384
+    -Wl,-z,common-page-size=16384)

package/android/build.gradle

@@ -13,7 +13,7 @@ def findNodeBinary() {
             return nodePath
         }
     }
-    
+
     // Check common node installation paths
     def homeDir = System.getProperty("user.home")
     def nodePaths = [
@@ -29,7 +29,7 @@ def findNodeBinary() {
             return path
         }
     }
-    
+
     // Fallback to 'node' (works if PATH is set correctly in terminal builds)
     return "node"
 }
@@ -82,10 +82,14 @@ def getExtOrIntegerDefault(name) {
 }
 
 // =============================================================================
-// Version Constants (MUST match Swift Package.swift and iOS Podspec)
+// Version Constants — read from sdk/runanywhere-commons/VERSION (single source of truth)
 // RABackendLlamaCPP from runanywhere-sdks
 // =============================================================================
-def coreVersion = "0.1.4"
+// Resolve through canonicalPath so the relative lookup survives the node_modules
+// symlink (node_modules/@runanywhere/llamacpp -> ../../sdk/runanywhere-react-native/packages/llamacpp);
+// fall back to the pinned version if the commons VERSION file can't be located.
+def coreVersionFile = file("${projectDir.canonicalPath}/../../../../runanywhere-commons/VERSION")
+def coreVersion = coreVersionFile.exists() ? coreVersionFile.text.trim() : "0.19.13"
 
 // =============================================================================
 // Binary Source - RABackendLlamaCPP from runanywhere-sdks
@@ -94,15 +98,63 @@ def githubOrg = "RunanywhereAI"
 def coreRepo = "runanywhere-sdks"
 
 // =============================================================================
-// testLocal Toggle
+// useLocalNatives Toggle (canonical name; testLocal kept as legacy alias)
 // =============================================================================
-def useLocalBuild = project.findProperty("runanywhere.testLocal")?.toBoolean() ?:
+def useLocalBuild = project.findProperty("runanywhere.useLocalNatives")?.toBoolean() ?:
+                    project.findProperty("runanywhere.testLocal")?.toBoolean() ?:
                     System.getenv("RA_TEST_LOCAL") == "1" ?: false
 
 // Native libraries directory
 def jniLibsDir = file("src/main/jniLibs")
 def downloadedLibsDir = file("build/downloaded-libs")
 
+def androidNdkHomeForRuntime() {
+    def explicitNdk = System.getenv("ANDROID_NDK_HOME") ?: System.getenv("NDK_HOME")
+    if (explicitNdk) {
+        return file(explicitNdk)
+    }
+
+    def androidSdk = System.getenv("ANDROID_HOME") ?:
+        System.getenv("ANDROID_SDK_ROOT") ?:
+        "${System.getProperty('user.home')}/Library/Android/sdk"
+    def ndkVersion = rootProject.ext.has("ndkVersion") ? rootProject.ext.ndkVersion : "27.0.12077973"
+    return file("${androidSdk}/ndk/${ndkVersion}")
+}
+
+def androidNdkHostTag() {
+    def osName = System.getProperty("os.name").toLowerCase(Locale.ROOT)
+    if (osName.contains("mac")) return "darwin-x86_64"
+    if (osName.contains("linux")) return "linux-x86_64"
+    throw new GradleException("Unsupported host for Android NDK runtime lookup: ${System.getProperty('os.name')}")
+}
+
+def androidNdkTripleForAbi(String abi) {
+    switch (abi) {
+        case 'arm64-v8a': return 'aarch64-linux-android'
+        case 'armeabi-v7a': return 'arm-linux-androideabi'
+        case 'x86_64': return 'x86_64-linux-android'
+        default: throw new GradleException("Unsupported Android ABI for libc++ runtime lookup: ${abi}")
+    }
+}
+
+def syncAndroidNdkRuntimeLibs(File outputDir) {
+    def ndkHome = androidNdkHomeForRuntime()
+    def prebuiltDir = new File(ndkHome, "toolchains/llvm/prebuilt/${androidNdkHostTag()}")
+    if (!prebuiltDir.isDirectory()) {
+        throw new GradleException("Android NDK prebuilt directory not found: ${prebuiltDir}")
+    }
+
+    ['arm64-v8a', 'armeabi-v7a', 'x86_64'].each { abi ->
+        def abiDir = new File(outputDir, abi)
+        def hasNativeLibs = abiDir.isDirectory() && abiDir.listFiles()?.any { it.name.endsWith(".so") }
+        if (!hasNativeLibs) return
+
+        def libcxx = new File(prebuiltDir, "sysroot/usr/lib/${androidNdkTripleForAbi(abi)}/libc++_shared.so")
+        if (!libcxx.isFile()) throw new GradleException("libc++_shared.so not found for ${abi} at ${libcxx}")
+        copy { from libcxx; into abiDir }
+    }
+}
+
 android {
     namespace "com.margelo.nitro.runanywhere.llama"
 
@@ -113,7 +165,7 @@ android {
         targetSdkVersion getExtOrIntegerDefault('targetSdkVersion')
 
         ndk {
-            abiFilters(*reactNativeArchitectures())
+            abiFilters 'arm64-v8a'
         }
 
         externalNativeBuild {
@@ -122,7 +174,7 @@ android {
                 arguments "-DANDROID_STL=c++_shared",
                           // Fix NitroModules prefab path - use app's build directory
                           "-DREACT_NATIVE_NITRO_BUILD_DIR=${rootProject.buildDir}"
-                abiFilters(*reactNativeArchitectures())
+                abiFilters 'arm64-v8a'
             }
         }
     }
@@ -130,6 +182,7 @@ android {
     externalNativeBuild {
         cmake {
             path "CMakeLists.txt"
+            version "3.30.5"
         }
     }
 
@@ -211,6 +264,12 @@ task downloadNativeLibs {
             return
         }
 
+        fileTree(dir: jniLibsDir, include: "**/libc++_shared.so").files.each { staleStdlib ->
+            if (staleStdlib.delete()) {
+                logger.lifecycle("[RunAnywhereLlama] Removed bundled libc++_shared.so: ${staleStdlib}")
+            }
+        }
+
         // Check if libs are already bundled for ALL requested ABIs (npm install case)
         def requestedAbis = reactNativeArchitectures()
         def allAbisBundled = requestedAbis.every { abi ->
@@ -382,6 +441,20 @@ task downloadNativeLibs {
     }
 }
 
+task syncAndroidRuntimeLibs {
+    description = "Stages 16 KB-aligned Android NDK libc++"
+    group = "build setup"
+
+    if (!useLocalBuild) {
+        dependsOn downloadNativeLibs
+    }
+
+    doLast {
+        syncAndroidNdkRuntimeLibs(jniLibsDir)
+        logger.lifecycle("[RunAnywhereLlama] Synced 16 KB-aligned Android NDK libc++ into ${jniLibsDir}")
+    }
+}
+
 if (!useLocalBuild) {
     preBuild.dependsOn downloadNativeLibs
 
@@ -394,6 +467,8 @@ if (!useLocalBuild) {
     }
 }
 
+preBuild.dependsOn syncAndroidRuntimeLibs
+
 // NOTE: cleanNativeLibs is NOT attached to clean task because npm-bundled libs should persist
 // Only use this task manually when needed during development
 task cleanNativeLibs(type: Delete) {

package/cpp/HybridRunAnywhereLlama.cpp

@@ -3,98 +3,42 @@
  *
  * Nitrogen HybridObject implementation for RunAnywhere Llama backend.
  *
- * Llama-specific implementation for text generation using LlamaCPP.
+ * Llama-specific provider registration for LlamaCPP.
  *
  * NOTE: LlamaCPP backend is REQUIRED and always linked via the build system.
  */
 
 #include "HybridRunAnywhereLlama.hpp"
 
-// Llama bridges
-#include "bridges/LLMBridge.hpp"
-#include "bridges/StructuredOutputBridge.hpp"
-
-// Backend registration header - always available
+// Backend registration headers - always available.
+// Resolve the canonical commons header (RACommons xcframework / jniLibs) rather
+// than a per-package fork, so the full published LlamaCPP ABI stays in sync.
 extern "C" {
-#include "rac_llm_llamacpp.h"
+#include "rac/backends/rac_llm_llamacpp.h"
 }
 
+#include "rac/core/rac_error.h"
+
 // Unified logging via rac_logger.h
 #include "rac_logger.h"
 
-#include <sstream>
-#include <chrono>
-#include <vector>
 #include <stdexcept>
+#include <string>
 
 // Log category for this module
 #define LOG_CATEGORY "LLM.LlamaCpp"
 
 namespace margelo::nitro::runanywhere::llama {
 
-using namespace ::runanywhere::bridges;
-
-// ============================================================================
-// JSON Utilities
-// ============================================================================
-
 namespace {
 
-int extractIntValue(const std::string& json, const std::string& key, int defaultValue) {
-  std::string searchKey = "\"" + key + "\":";
-  size_t pos = json.find(searchKey);
-  if (pos == std::string::npos) return defaultValue;
-  pos += searchKey.length();
-  while (pos < json.size() && (json[pos] == ' ' || json[pos] == '\t')) pos++;
-  if (pos >= json.size()) return defaultValue;
-  return std::stoi(json.substr(pos));
-}
-
-float extractFloatValue(const std::string& json, const std::string& key, float defaultValue) {
-  std::string searchKey = "\"" + key + "\":";
-  size_t pos = json.find(searchKey);
-  if (pos == std::string::npos) return defaultValue;
-  pos += searchKey.length();
-  while (pos < json.size() && (json[pos] == ' ' || json[pos] == '\t')) pos++;
-  if (pos >= json.size()) return defaultValue;
-  return std::stof(json.substr(pos));
-}
-
-std::string extractStringValue(const std::string& json, const std::string& key, const std::string& defaultValue = "") {
-  std::string searchKey = "\"" + key + "\":\"";
-  size_t pos = json.find(searchKey);
-  if (pos == std::string::npos) return defaultValue;
-  pos += searchKey.length();
-  size_t endPos = json.find("\"", pos);
-  if (endPos == std::string::npos) return defaultValue;
-  return json.substr(pos, endPos - pos);
-}
-
-std::string buildJsonObject(const std::vector<std::pair<std::string, std::string>>& keyValues) {
-  std::string result = "{";
-  for (size_t i = 0; i < keyValues.size(); i++) {
-    if (i > 0) result += ",";
-    result += "\"" + keyValues[i].first + "\":" + keyValues[i].second;
-  }
-  result += "}";
-  return result;
-}
-
-std::string jsonString(const std::string& value) {
-  std::string escaped = "\"";
-  for (char c : value) {
-    if (c == '"') escaped += "\\\"";
-    else if (c == '\\') escaped += "\\\\";
-    else if (c == '\n') escaped += "\\n";
-    else if (c == '\r') escaped += "\\r";
-    else if (c == '\t') escaped += "\\t";
-    else escaped += c;
-  }
-  escaped += "\"";
-  return escaped;
+bool isRegistrationSuccess(rac_result_t result) {
+  return result == RAC_SUCCESS ||
+         result == RAC_ERROR_MODULE_ALREADY_REGISTERED ||
+         result == RAC_ERROR_PLUGIN_DUPLICATE;
 }
 
-} // anonymous namespace
+} // namespace
 
 // ============================================================================
 // Constructor / Destructor
@@ -106,7 +50,6 @@ HybridRunAnywhereLlama::HybridRunAnywhereLlama() : HybridObject(TAG) {
 
 HybridRunAnywhereLlama::~HybridRunAnywhereLlama() {
   RAC_LOG_DEBUG(LOG_CATEGORY, "HybridRunAnywhereLlama destructor");
-  LLMBridge::shared().destroy();
 }
 
 // ============================================================================
@@ -117,17 +60,23 @@ std::shared_ptr<Promise<bool>> HybridRunAnywhereLlama::registerBackend() {
   return Promise<bool>::async([this]() {
     RAC_LOG_DEBUG(LOG_CATEGORY, "Registering LlamaCPP backend with C++ registry");
 
+    // rac_backend_llamacpp_register() internally registers both the module and
+    // the plugin entry (rac_plugin_entry_llamacpp()) — see
+    // engines/llamacpp/rac_backend_llamacpp_register.cpp:462-478, which plugs
+    // the Android dynamic-loading gap where the RAC_STATIC_PLUGIN_REGISTER ctor
+    // never fires (Kotlin/JNI loads librac_backend_llamacpp_jni.so directly,
+    // not the carrier librunanywhere_llamacpp.so). Matches the iOS Swift
+    // source-of-truth in sdk/runanywhere-swift/Sources/LlamaCPPRuntime which
+    // also relies on the in-engine registration only.
     rac_result_t result = rac_backend_llamacpp_register();
-    // RAC_SUCCESS (0) or RAC_ERROR_MODULE_ALREADY_REGISTERED (-4) are both OK
-    if (result == RAC_SUCCESS || result == -4) {
-      RAC_LOG_INFO(LOG_CATEGORY, "LlamaCPP backend registered successfully");
-      isRegistered_ = true;
-      return true;
-    } else {
+    if (!isRegistrationSuccess(result)) {
       RAC_LOG_ERROR(LOG_CATEGORY, "LlamaCPP registration failed with code: %d", result);
-      setLastError("LlamaCPP registration failed with error: " + std::to_string(result));
       throw std::runtime_error("LlamaCPP registration failed with error: " + std::to_string(result));
     }
+
+    RAC_LOG_INFO(LOG_CATEGORY, "LlamaCPP backend registered successfully");
+    isRegistered_ = true;
+    return true;
   });
 }
 
@@ -151,196 +100,4 @@ std::shared_ptr<Promise<bool>> HybridRunAnywhereLlama::isBackendRegistered() {
   });
 }
 
-// ============================================================================
-// Model Loading
-// ============================================================================
-
-std::shared_ptr<Promise<bool>> HybridRunAnywhereLlama::loadModel(
-    const std::string& path,
-    const std::optional<std::string>& modelId,
-    const std::optional<std::string>& modelName,
-    const std::optional<std::string>& configJson) {
-  return Promise<bool>::async([this, path, modelId, modelName, configJson]() {
-    std::lock_guard<std::mutex> lock(modelMutex_);
-
-    RAC_LOG_INFO(LOG_CATEGORY, "Loading Llama model: %s", path.c_str());
-
-    std::string id = modelId.value_or("");
-    std::string name = modelName.value_or("");
-
-    // Call with correct 4-arg signature (path, modelId, modelName)
-    // LLMBridge::loadModel will throw on error
-    auto result = LLMBridge::shared().loadModel(path, id, name);
-    if (result != 0) {
-      std::string error = "Failed to load Llama model: " + path + " (error: " + std::to_string(result) + ")";
-      setLastError(error);
-      throw std::runtime_error(error);
-    }
-    return true;
-  });
-}
-
-std::shared_ptr<Promise<bool>> HybridRunAnywhereLlama::isModelLoaded() {
-  return Promise<bool>::async([]() {
-    return LLMBridge::shared().isLoaded();
-  });
-}
-
-std::shared_ptr<Promise<bool>> HybridRunAnywhereLlama::unloadModel() {
-  return Promise<bool>::async([this]() {
-    std::lock_guard<std::mutex> lock(modelMutex_);
-    auto result = LLMBridge::shared().unload();
-    return result == 0;
-  });
-}
-
-std::shared_ptr<Promise<std::string>> HybridRunAnywhereLlama::getModelInfo() {
-  return Promise<std::string>::async([]() {
-    if (!LLMBridge::shared().isLoaded()) {
-      return std::string("{}");
-    }
-    return buildJsonObject({
-      {"loaded", "true"},
-      {"backend", jsonString("llamacpp")}
-    });
-  });
-}
-
-// ============================================================================
-// Text Generation
-// ============================================================================
-
-std::shared_ptr<Promise<std::string>> HybridRunAnywhereLlama::generate(
-    const std::string& prompt,
-    const std::optional<std::string>& optionsJson) {
-  return Promise<std::string>::async([this, prompt, optionsJson]() {
-    if (!LLMBridge::shared().isLoaded()) {
-      setLastError("Model not loaded");
-      throw std::runtime_error("LLMBridge: Model not loaded. Call loadModel() first.");
-    }
-
-    LLMOptions options;
-    if (optionsJson.has_value()) {
-      options.maxTokens = extractIntValue(*optionsJson, "max_tokens", 512);
-      options.temperature = extractFloatValue(*optionsJson, "temperature", 0.7f);
-      options.topP = extractFloatValue(*optionsJson, "top_p", 0.9f);
-      options.topK = extractIntValue(*optionsJson, "top_k", 40);
-    }
-
-    RAC_LOG_DEBUG(LOG_CATEGORY, "Generating with prompt: %.50s...", prompt.c_str());
-
-    auto startTime = std::chrono::high_resolution_clock::now();
-    // LLMBridge::generate will throw on error
-    auto result = LLMBridge::shared().generate(prompt, options);
-    auto endTime = std::chrono::high_resolution_clock::now();
-    auto durationMs = std::chrono::duration_cast<std::chrono::milliseconds>(
-        endTime - startTime).count();
-
-    return buildJsonObject({
-      {"text", jsonString(result.text)},
-      {"tokensUsed", std::to_string(result.tokenCount)},
-      {"latencyMs", std::to_string(durationMs)},
-      {"cancelled", result.cancelled ? "true" : "false"}
-    });
-  });
-}
-
-std::shared_ptr<Promise<std::string>> HybridRunAnywhereLlama::generateStream(
-    const std::string& prompt,
-    const std::string& optionsJson,
-    const std::function<void(const std::string&, bool)>& callback) {
-  return Promise<std::string>::async([this, prompt, optionsJson, callback]() {
-    if (!LLMBridge::shared().isLoaded()) {
-      setLastError("Model not loaded");
-      throw std::runtime_error("LLMBridge: Model not loaded. Call loadModel() first.");
-    }
-
-    LLMOptions options;
-    options.maxTokens = extractIntValue(optionsJson, "max_tokens", 512);
-    options.temperature = extractFloatValue(optionsJson, "temperature", 0.7f);
-
-    std::string fullResponse;
-    std::string streamError;
-
-    LLMStreamCallbacks streamCallbacks;
-    streamCallbacks.onToken = [&callback, &fullResponse](const std::string& token) -> bool {
-      fullResponse += token;
-      if (callback) {
-        callback(token, false);
-      }
-      return true;
-    };
-    streamCallbacks.onComplete = [&callback](const std::string&, int, double) {
-      if (callback) {
-        callback("", true);
-      }
-    };
-    streamCallbacks.onError = [this, &streamError](int code, const std::string& message) {
-      setLastError(message);
-      streamError = message;
-    };
-
-    LLMBridge::shared().generateStream(prompt, options, streamCallbacks);
-
-    if (!streamError.empty()) {
-      throw std::runtime_error("LLMBridge: Stream generation failed: " + streamError);
-    }
-
-    return fullResponse;
-  });
-}
-
-std::shared_ptr<Promise<bool>> HybridRunAnywhereLlama::cancelGeneration() {
-  return Promise<bool>::async([]() {
-    LLMBridge::shared().cancel();
-    return true;
-  });
-}
-
-// ============================================================================
-// Structured Output
-// ============================================================================
-
-std::shared_ptr<Promise<std::string>> HybridRunAnywhereLlama::generateStructured(
-    const std::string& prompt,
-    const std::string& schema,
-    const std::optional<std::string>& optionsJson) {
-  return Promise<std::string>::async([this, prompt, schema, optionsJson]() {
-    auto result = StructuredOutputBridge::shared().generate(
-      prompt, schema, optionsJson.value_or("")
-    );
-
-    if (result.success) {
-      return result.json;
-    } else {
-      setLastError(result.error);
-      return buildJsonObject({{"error", jsonString(result.error)}});
-    }
-  });
-}
-
-// ============================================================================
-// Utilities
-// ============================================================================
-
-std::shared_ptr<Promise<std::string>> HybridRunAnywhereLlama::getLastError() {
-  return Promise<std::string>::async([this]() { return lastError_; });
-}
-
-std::shared_ptr<Promise<double>> HybridRunAnywhereLlama::getMemoryUsage() {
-  return Promise<double>::async([]() {
-    // TODO: Get memory usage from LlamaCPP
-    return 0.0;
-  });
-}
-
-// ============================================================================
-// Helper Methods
-// ============================================================================
-
-void HybridRunAnywhereLlama::setLastError(const std::string& error) {
-  lastError_ = error;
-  RAC_LOG_ERROR(LOG_CATEGORY, "Error: %s", error.c_str());
-}
-
 } // namespace margelo::nitro::runanywhere::llama