Unable to build a dll using libtorch + cuda

I am trying to build a shared library to then call it from python scripts using ctypes. I had everything working with no cuda support, but did not manage to make it work on the GPU. Below is a minimal example:

I downloaded libtorch-win-shared-with-deps-1.13.1+cu117 from pytorch.org and unzipped it to C: I then used the following cmake files to build both an executable and a shared library:

cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
list(APPEND CMAKE_PREFIX_PATH "C:\\libtorch")
project(example-app)

find_package(Torch REQUIRED)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")

add_library(example-lib SHARED example-app.cpp example-app.h)
target_link_libraries(example-lib "${TORCH_LIBRARIES}")
set_property(TARGET example-lib PROPERTY CXX_STANDARD 14)

add_executable(example-app example-app.cpp example-app.h)
target_link_libraries(example-app "${TORCH_LIBRARIES}")
set_property(TARGET example-app PROPERTY CXX_STANDARD 14)

if (MSVC)
  file(GLOB TORCH_DLLS "${TORCH_INSTALL_PREFIX}/lib/*.dll")
  add_custom_command(TARGET example-app
                     POST_BUILD
                     COMMAND ${CMAKE_COMMAND} -E copy_if_different
                     ${TORCH_DLLS}
                     $<TARGET_FILE_DIR:example-app>)
endif (MSVC)

The example-app.cpp script is

#include <torch/torch.h>
#include <iostream>

#include "example-app.h"

int test() {
  torch::Tensor tensor = torch::rand({3, 2});
  std::cout << tensor << std::endl;
  return 0;
}

int main() {
	test();
	return 0;
}

and the header

extern "C" { __declspec(dllexport) int test(); }

I then want to call test in a python script with

import os
import pathlib
import ctypes

libname = pathlib.Path("C:/Users/Matrix/PycharmProjects/torch_test/buil/Release")
os.add_dll_directory(str(libname))
c_lib = ctypes.CDLL(str(libname / "example-lib.dll"), winmode=0)

print(c_lib.test())

But I get the following error Procedure entry point could not be located in the dynamic link library
(sorry for the french):
Screenshot from 2022-12-16 15-52-05

Note that calling the executable works as desired. Any idea what could be causing the problem?
Configurations: Windows 10, MSVC, CUDA 11.7
Thanks for any help