Install required runtime dlls to package directory on windows

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.4...3.22)
+cmake_minimum_required(VERSION 3.21)
 
 project(llama_cpp)
 
@@ -33,4 +33,13 @@ if (LLAMA_BUILD)
         FRAMEWORK DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp
         RESOURCE DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp
     )
+    # Workaround for Windows + CUDA https://github.com/abetlen/llama-cpp-python/issues/563
+    install(
+        FILES $<TARGET_RUNTIME_DLLS:llama>
+        DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp
+    )
+    install(
+        FILES $<TARGET_RUNTIME_DLLS:llama>
+        DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp
+    )
 endif()

llama_cpp/llama_cpp.py

@@ -58,7 +58,7 @@ def _load_shared_library(lib_base_name: str):
         if "CUDA_PATH" in os.environ:
             os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin"))
             os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib"))
-        cdll_args["winmode"] = 0
+        cdll_args["winmode"] = ctypes.RTLD_GLOBAL
 
     # Try to load the shared library, handling potential errors
     for _lib_path in _lib_paths:

pyproject.toml

@@ -54,7 +54,7 @@ all = [
 [tool.scikit-build]
 wheel.packages = ["llama_cpp"]
 cmake.verbose = true
-cmake.minimum-version = "3.12"
+cmake.minimum-version = "3.21"
 minimum-version = "0.5"
 sdist.exclude = [".git", "vendor/llama.cpp/.git"]