fix: fix windows build

CMakeLists.txt

@@ -64,6 +64,17 @@ set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(LLAMA_STATIC ON CACHE BOOL "Build llama as static library")
 add_subdirectory("src/llama.cpp")
 
+# apply patches
+set(PATCH_FILE ${CMAKE_SOURCE_DIR}/patches/llama.patch)
+add_custom_target(patch)
+add_custom_command(
+  TARGET patch
+  COMMAND patch -N -p1 < ${PATCH_FILE} || true
+  WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/src/llama.cpp
+
+)
+add_dependencies(ggml patch)
+
 file(
   GLOB SOURCE_FILES
     "src/addons.cc"
@@ -95,6 +106,7 @@ if (LLAMA_QNN)
     message(FATAL_ERROR "QNN is not supported on this platform")
   endif()
   set(QNN_LIB_PATH ${QNN_ROOT}/lib/${QNN_PLATFORM})
+  message(STATUS "QNN_LIB_PATH: ${QNN_LIB_PATH}")
 
   file(
     GLOB QNN_SO_FILES
@@ -118,33 +130,22 @@ if (LLAMA_QNN)
 
   file(
     GLOB QNN_HEADER_FILES
-      "src/ggml-qnn.h"
+      "src/ggml-qnn/ggml-qnn.h"
   )
 
   file(
     GLOB QNN_SOURCE_FILES
-      "src/ggml-qnn.cpp"
+      "src/ggml-qnn/pthread-shim.h"
+      "src/ggml-qnn/ggml-qnn.cpp"
   )
 
   target_compile_definitions(ggml PUBLIC GGML_USE_QNN)
-  target_include_directories(ggml PUBLIC ${QNN_ROOT}/include/QNN)
+  target_include_directories(ggml PUBLIC ${QNN_ROOT}/include ${QNN_ROOT}/include/QNN)
   target_sources(ggml PRIVATE ${QNN_SOURCE_FILES} ${QNN_HEADER_FILES})
-  target_include_directories(llama PRIVATE "src")
+  target_include_directories(llama PRIVATE "src/ggml-qnn")
   set_target_properties(ggml PROPERTIES CXX_STANDARD 17)
-
-  # apply patches/qnn.patch to ggml
-  add_custom_command(
-    OUTPUT ${CMAKE_BUILD_DIR}/patch.log
-    COMMAND git apply ${CMAKE_SOURCE_DIR}/patches/qnn.patch
-    WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/src/llama.cpp
-  )
-else()
-  # undo patches/qnn.patch to ggml
-  add_custom_command(
-    OUTPUT ${CMAKE_BUILD_DIR}/patch.log
-    COMMAND git apply -R ${CMAKE_SOURCE_DIR}/patches/qnn.patch
-    WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/src/llama.cpp
-  )
+  set_target_properties(ggml PROPERTIES CXX_STANDARD_REQUIRED ON)
+  set_target_properties(ggml PROPERTIES C_STANDARD 11)
 endif()
 
 add_library(${PROJECT_NAME} SHARED ${SOURCE_FILES} ${CMAKE_JS_SRC})

package.json

@@ -39,10 +39,8 @@
   },
   "files": [
     "bin/**/*",
-    "scripts/*.js",
-    "scripts/*.ts",
-    "src/*",
-    "externals/**/*.{c,cc,cpp,h,hh,hpp,txt,cmake}",
+    "patches/*",
+    "src/**/*.{c,cc,cpp,h,hh,hpp,txt,cmake}",
     "lib/*.js",
     "lib/*.ts",
     "CMakeLists.txt"

patches/qnn.patch → patches/llama.patch

@@ -1,16 +1,7 @@
 diff --git a/ggml-backend.c b/ggml-backend.c
-index f5bdcf07..536a5767 100644
+index e91d97cd..be4989d3 100644
 --- a/ggml-backend.c
 +++ b/ggml-backend.c
-@@ -416,7 +416,7 @@ GGML_CALL static void ggml_backend_registry_init(void) {
-     }
-
-     initialized = true;
--
-+    printf("GGML_USE_CPU\n");
-     ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL);
-
-     // add forward decls here to avoid including the backend headers
 @@ -445,6 +445,10 @@ GGML_CALL static void ggml_backend_registry_init(void) {
      extern GGML_CALL void ggml_backend_kompute_reg_devices(void);
      ggml_backend_kompute_reg_devices();
@@ -23,7 +14,7 @@ index f5bdcf07..536a5767 100644
 
  GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
 diff --git a/llama.cpp b/llama.cpp
-index 18d6297c..f2a39613 100644
+index a25d115c..ff0d929f 100644
 --- a/llama.cpp
 +++ b/llama.cpp
 @@ -17,6 +17,8 @@
@@ -35,7 +26,7 @@ index 18d6297c..f2a39613 100644
  #endif
 
  #ifdef GGML_USE_METAL
-@@ -1679,6 +1681,8 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) {
+@@ -1658,6 +1660,8 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) {
      buft = ggml_backend_opencl_buffer_type();
  #elif defined(GGML_USE_KOMPUTE)
      buft = ggml_backend_kompute_buffer_type(gpu);
@@ -44,18 +35,16 @@ index 18d6297c..f2a39613 100644
      if (buft == nullptr) {
          LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu);
      }
-@@ -15293,8 +15297,9 @@ bool llama_supports_mlock(void) {
+@@ -14916,7 +14920,7 @@ bool llama_supports_mlock(void) {
 
  bool llama_supports_gpu_offload(void) {
  #if defined(GGML_USE_CUDA) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
 -    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
 +    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_QNN)
      // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
-+    printf("llama_supports_gpu_offload: true\n");
      return true;
  #else
-     return false;
-@@ -15607,6 +15612,16 @@ struct llama_context * llama_new_context_with_model(
+@@ -15203,6 +15207,16 @@ struct llama_context * llama_new_context_with_model(
              }
              ctx->backends.push_back(backend);
          }