@novastera-oss/llamarn 0.4.0 → 0.4.1

package/RNLlamaCpp.podspec

@@ -53,7 +53,7 @@ Pod::Spec.new do |s|
   # Compiler settings
   s.pod_target_xcconfig = {
     "HEADER_SEARCH_PATHS" => "\"$(PODS_TARGET_SRCROOT)/ios/include\" \"$(PODS_TARGET_SRCROOT)/cpp\" \"$(PODS_TARGET_SRCROOT)/ios/generated/RNLlamaCppSpec\" \"$(PODS_TARGET_SRCROOT)/ios/generated\" \"$(PODS_TARGET_SRCROOT)/cpp/llama.cpp\" \"$(PODS_TARGET_SRCROOT)/cpp/llama.cpp/include\" \"$(PODS_TARGET_SRCROOT)/cpp/llama.cpp/ggml/include\" \"$(PODS_TARGET_SRCROOT)/cpp/llama.cpp/common\" \"$(PODS_TARGET_SRCROOT)/cpp/llama.cpp/vendor\" \"$(PODS_ROOT)/boost\" \"$(PODS_ROOT)/Headers/Public/React-bridging\" \"$(PODS_ROOT)/Headers/Public/React\"",
-    "OTHER_CPLUSPLUSFLAGS" => "-DFOLLY_NO_CONFIG -DFOLLY_MOBILE=1 -DFOLLY_USE_LIBCPP=1 -DLLAMA_METAL -DRCT_NEW_ARCH_ENABLED=1 -DFBJSRT_EXPORTED=1",
+    "OTHER_CPLUSPLUSFLAGS" => "-DFOLLY_NO_CONFIG -DFOLLY_MOBILE=1 -DFOLLY_USE_LIBCPP=1 -DFOLLY_CFG_NO_COROUTINES=1 -DLLAMA_METAL -DRCT_NEW_ARCH_ENABLED=1 -DFBJSRT_EXPORTED=1",
     "CLANG_CXX_LANGUAGE_STANDARD" => "c++17",
     "GCC_OPTIMIZATION_LEVEL" => "3", # Maximum optimization
     "SWIFT_OPTIMIZATION_LEVEL" => "-O",

package/android/CMakeLists.txt

@@ -78,9 +78,17 @@ add_library(
     ${CPP_DIR}/rn-completion.cpp
 )
 
-# Suppress unused function warnings for llama.cpp code
-target_compile_options(common PRIVATE -Wno-unused-function)
-target_compile_options(RNLlamaCpp PRIVATE -Wno-unused-function)
+# Suppress additional warnings that are treated as errors in Expo SDK 54
+target_compile_options(common PRIVATE )
+
+# Use React Native's compile options function for proper C++ flags and RN_SERIALIZABLE_STATE
+if(ReactAndroid_VERSION_MINOR GREATER_EQUAL 80)
+    # Add additional warning suppressions for RNLlamaCpp target
+    target_compile_reactnative_options(RNLlamaCpp PRIVATE)
+    target_compile_options(RNLlamaCpp PRIVATE -Wno-unused-function)
+else()
+    target_compile_options(RNLlamaCpp PRIVATE -Wno-unused-function)
+endif()
 
 # Check if Vulkan backend library is available
 set(VULKAN_BACKEND_AVAILABLE FALSE)

package/cpp/LlamaCppModel.cpp

@@ -948,16 +948,8 @@ jsi::Value LlamaCppModel::embeddingJsi(jsi::Runtime& rt, const jsi::Value* args,
       throw std::runtime_error("Invalid embedding dimension");
     }
 
-    // For OpenAI compatibility, default to mean pooling
-    enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_MEAN;
-    if (options.hasProperty(rt, "pooling") && options.getProperty(rt, "pooling").isString()) {
-      std::string pooling = options.getProperty(rt, "pooling").getString(rt).utf8(rt);
-      if (pooling == "last") {
-        pooling_type = LLAMA_POOLING_TYPE_LAST;
-      } else if (pooling == "cls" || pooling == "first") {
-        pooling_type = LLAMA_POOLING_TYPE_CLS;
-      }
-    }
+    // Note: Pooling is handled automatically by llama_get_embeddings()
+    // The function returns the appropriate embedding based on the model's configuration
 
     // Get the embeddings
     std::vector<float> embedding_vec(n_embd);

package/cpp/SystemUtils.cpp

@@ -20,10 +20,6 @@
 
 namespace facebook::react {
 
-// Memory fallback constants (clearly defined for future maintenance)
-constexpr int64_t FALLBACK_IOS_MEMORY = 2LL * 1024 * 1024 * 1024;     // 2GB default
-constexpr int64_t FALLBACK_ANDROID_MEMORY = 3LL * 1024 * 1024 * 1024; // 3GB for Android
-constexpr int64_t DEFAULT_FALLBACK_MEMORY = 2LL * 1024 * 1024 * 1024; // 2GB default
 
 int SystemUtils::getOptimalThreadCount() {
     int cpuCores = std::thread::hardware_concurrency();
@@ -85,11 +81,11 @@ int64_t getTotalPhysicalMemory() {
     // Fallback to a conservative estimate if we couldn't get the actual memory
     if (total_memory <= 0) {
 #if defined(__APPLE__) && TARGET_OS_IPHONE
-        total_memory = FALLBACK_IOS_MEMORY;
+        total_memory = 2LL * 1024 * 1024 * 1024;     // 2GB default for iOS
 #elif defined(__ANDROID__)
-        total_memory = FALLBACK_ANDROID_MEMORY;
+        total_memory = 3LL * 1024 * 1024 * 1024;     // 3GB default for Android
 #else
-        total_memory = DEFAULT_FALLBACK_MEMORY;
+        total_memory = 2LL * 1024 * 1024 * 1024;     // 2GB default for other platforms
 #endif
     }
 

package/cpp/rn-completion.cpp

@@ -255,7 +255,7 @@ CompletionResult run_completion(
         }
 
         // Start generating tokens
-        const int64_t t_start_generation = ggml_time_us();
+        // Note: Timing variables removed as they were not being used
 
         while (state.has_next_token && state.n_remaining > 0) {
             // Sample the next token
@@ -323,9 +323,7 @@ CompletionResult run_completion(
             }
         }
 
-        const int64_t t_end_generation = ggml_time_us();
-        // Note: keeping generation_time_ms for future timing measurements
-        // const double generation_time_ms = (t_end_generation - t_start_generation) / 1000.0;
+        // Note: Timing measurements removed as they were not being used
 
         // Set the result
         result.content = state.generated_text;
@@ -428,27 +426,7 @@ CompletionResult run_chat_completion(
             // Default to grammar_triggers provided by chat_params
             cmpl_options.grammar_triggers = chat_params.grammar_triggers;
 
-            bool original_grammar_lazy = chat_params.grammar_lazy; // Store original for logging
-
-            // Add a debug log to observe final grammar_lazy and grammar_triggers
-            /*
-            if (callback) {
-                std::string tool_choice_str;
-                switch (template_inputs.tool_choice) {
-                    case COMMON_CHAT_TOOL_CHOICE_AUTO: tool_choice_str = "auto"; break;
-                    case COMMON_CHAT_TOOL_CHOICE_NONE: tool_choice_str = "none"; break;
-                    case COMMON_CHAT_TOOL_CHOICE_REQUIRED: tool_choice_str = "required"; break;
-                    default: tool_choice_str = "unknown"; break;
-                }
-                std::string debug_msg = "[DEBUG CHAT_PARAMS] grammar_lazy: " +
-                                      std::string(cmpl_options.grammar_lazy ? "true" : "false") +
-                                      " | grammar_triggers_count: " + std::to_string(cmpl_options.grammar_triggers.size()) + // Log triggers from cmpl_options
-                                      " | For Tool Choice: " + tool_choice_str +
-                                      " | Parallel Tool Calls: " + std::string(template_inputs.parallel_tool_calls ? "true" : "false") +
-                                      " | Original chat_params.grammar_lazy: " + std::string(original_grammar_lazy ? "true" : "false"); // Log original lazy
-                callback(debug_msg, false);
-            }
-            */
+            // Note: Debug logging removed as it was not being used
         }
 
         // Run standard completion with the processed prompt
@@ -534,5 +512,3 @@ CompletionResult run_chat_completion(
 }
 
 } // namespace facebook::react
-
-

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@novastera-oss/llamarn",
-  "version": "0.4.0",
+  "version": "0.4.1",
   "description": "An attempt at a pure cpp turbo module library",
   "source": "./src/index.tsx",
   "main": "./lib/module/index.js",