@novastera-oss/llamarn 0.1.3-beta.3 → 0.1.3-beta.5

package/android/CMakeLists.txt

@@ -74,22 +74,59 @@ add_library(
 target_compile_options(common PRIVATE -Wno-unused-function)
 target_compile_options(RNLlamaCpp PRIVATE -Wno-unused-function)
 
+# Check if Vulkan backend library is available
+set(VULKAN_BACKEND_AVAILABLE FALSE)
+if(EXISTS ${JNI_LIBS_DIR}/${ANDROID_ABI}/libggml-vulkan.so)
+    set(VULKAN_BACKEND_AVAILABLE TRUE)
+    message(STATUS "Vulkan backend library found for ${ANDROID_ABI}")
+else()
+    message(STATUS "Vulkan backend library not found for ${ANDROID_ABI}")
+endif()
+
+# Check if OpenCL backend library is available
+set(OPENCL_BACKEND_AVAILABLE FALSE)
+if(EXISTS ${JNI_LIBS_DIR}/${ANDROID_ABI}/libggml-opencl.so AND EXISTS ${JNI_LIBS_DIR}/${ANDROID_ABI}/libOpenCL.so)
+    set(OPENCL_BACKEND_AVAILABLE TRUE)
+    message(STATUS "OpenCL backend libraries found for ${ANDROID_ABI}")
+else()
+    message(STATUS "OpenCL backend libraries not found for ${ANDROID_ABI}")
+endif()
+
 # Hybrid backend approach: CPU static (built into main libraries), GPU dynamic
 # CPU backend will be statically linked into main libraries (libggml.so, libllama.so)
-# GPU backends (OpenCL, Vulkan) will be dynamically loaded at runtime
+# GPU backends (OpenCL, Vulkan) will be dynamically loaded at runtime only if available
 target_compile_definitions(common PRIVATE 
     -DGGML_BACKEND_DL=1  # Enable dynamic loading for GPU backends
     -DGGML_CPU=1         # CPU backend statically built into main libraries
-    -DGGML_VULKAN=1      # Enable Vulkan backend support for dynamic loading
-    -DGGML_OPENCL=1      # Enable OpenCL backend support for dynamic loading
 )
 target_compile_definitions(RNLlamaCpp PRIVATE 
     -DGGML_BACKEND_DL=1  # Enable dynamic loading for GPU backends
     -DGGML_CPU=1         # CPU backend statically built into main libraries
-    -DGGML_VULKAN=1      # Enable Vulkan backend support for dynamic loading
-    -DGGML_OPENCL=1      # Enable OpenCL backend support for dynamic loading
 )
 
+# DISABLE Vulkan on Android - causes crashes during auto-initialization on emulators
+# Even with n_gpu_layers=0, llama.cpp tries to initialize all available backends
+message(STATUS "Vulkan backend support DISABLED on Android to prevent emulator crashes")
+
+# TODO: Enable Vulkan backend if available (currently disabled due to emulator crashes)
+# Uncomment the lines below to test Vulkan support on real devices
+# if(VULKAN_BACKEND_AVAILABLE)
+#     target_compile_definitions(common PRIVATE -DGGML_VULKAN=1)
+#     target_compile_definitions(RNLlamaCpp PRIVATE -DGGML_VULKAN=1)
+#     message(STATUS "Vulkan backend support enabled for dynamic loading")
+# else()
+#     message(STATUS "Vulkan backend support disabled - library not available")
+# endif()
+
+# Enable OpenCL backend if available
+if(OPENCL_BACKEND_AVAILABLE)
+    target_compile_definitions(common PRIVATE -DGGML_OPENCL=1)
+    target_compile_definitions(RNLlamaCpp PRIVATE -DGGML_OPENCL=1)
+    message(STATUS "OpenCL backend support enabled for dynamic loading")
+else()
+    message(STATUS "OpenCL backend support disabled - library not available")
+endif()
+
 # Include directories
 target_include_directories(common PRIVATE
     ${CPP_DIR}
@@ -161,12 +198,18 @@ add_custom_command(TARGET RNLlamaCpp POST_BUILD
 
 # Also copy any optional GPU libraries if they exist
 if(EXISTS ${JNI_LIBS_DIR}/${ANDROID_ABI}/libggml-vulkan.so)
-    add_custom_command(TARGET RNLlamaCpp POST_BUILD
-        COMMAND ${CMAKE_COMMAND} -E copy_if_different
-            ${JNI_LIBS_DIR}/${ANDROID_ABI}/libggml-vulkan.so
-            $<TARGET_FILE_DIR:RNLlamaCpp>/libggml-vulkan.so
-        COMMENT "Copying Vulkan library to build output directory"
-    )
+    # Don't copy Vulkan backend on Android - it crashes on emulators during auto-initialization
+    # Even with n_gpu_layers=0, llama.cpp tries to initialize all available backends
+    # and the Android emulator Vulkan driver is broken
+    message(STATUS "Skipping Vulkan backend copy to prevent emulator crashes")
+    
+    # TODO: Uncomment the lines below to enable Vulkan library copying for testing on real devices
+    # add_custom_command(TARGET RNLlamaCpp POST_BUILD
+    #     COMMAND ${CMAKE_COMMAND} -E copy_if_different
+    #         ${JNI_LIBS_DIR}/${ANDROID_ABI}/libggml-vulkan.so
+    #         $<TARGET_FILE_DIR:RNLlamaCpp>/libggml-vulkan.so
+    #     COMMENT "Copying Vulkan backend library to build output directory"
+    # )
 endif()
 
 if(EXISTS ${JNI_LIBS_DIR}/${ANDROID_ABI}/libggml-opencl.so)

package/cpp/PureCppImpl.cpp

@@ -245,40 +245,8 @@ jsi::Value PureCppImpl::initLlama(jsi::Runtime &runtime, jsi::Object options) {
         throw std::runtime_error("Failed to initialize model and context");
       }
     } catch (const std::exception& e) {
-      std::string error_msg = e.what();
-      
-      // Check for specific Vulkan errors that indicate Adreno GPU shader incompatibility
-      bool isVulkanShaderError = (
-        error_msg.find("createComputePipeline") != std::string::npos ||
-        error_msg.find("ErrorUnknown") != std::string::npos ||
-        error_msg.find("matmul_q4_k") != std::string::npos ||
-        error_msg.find("matmul_q5_k") != std::string::npos ||
-        error_msg.find("dequant_q4_K") != std::string::npos ||
-        error_msg.find("dequant_q5_K") != std::string::npos ||
-        error_msg.find("vulkan") != std::string::npos
-      );
-      
       // If we were trying to use GPU and got a Vulkan/shader error, retry with CPU-only
-      if (params.n_gpu_layers > 0 && isVulkanShaderError) {
-        fprintf(stderr, "Vulkan shader compilation failed (likely Adreno GPU incompatibility): %s\n", e.what());
-        fprintf(stderr, "This is a known issue with Q4_K/Q5_K shaders on Qualcomm Adreno GPUs\n");
-        fprintf(stderr, "Retrying with CPU-only mode...\n");
-        
-        // Retry with CPU-only
-        params.n_gpu_layers = 0;
-        
-        try {
-          result = common_init_from_params(params);
-          
-          if (!result.model || !result.context) {
-            throw std::runtime_error("Failed to initialize model and context even with CPU-only mode");
-          }
-          
-          fprintf(stderr, "Successfully recovered with CPU-only mode after Vulkan shader failure\n");
-        } catch (const std::exception& cpu_e) {
-          throw std::runtime_error(std::string("Model initialization failed: ") + cpu_e.what());
-        }
-      } else if (params.n_gpu_layers > 0) {
+      if (params.n_gpu_layers > 0) {
         // Other GPU error, still try CPU fallback
         fprintf(stderr, "GPU initialization failed (%s), retrying with CPU-only\n", e.what());
         

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@novastera-oss/llamarn",
-  "version": "0.1.3-beta.3",
+  "version": "0.1.3-beta.5",
   "description": "An attempt at a pure cpp turbo module library",
   "source": "./src/index.tsx",
   "main": "./lib/module/index.js",