Dynamic loading libpython with pybind11

I'm trying to build some shared library with pybind11 from Mac OSX. I'm running into the error:

dyld: Symbol not found: _PyBaseObject_Type
  Referenced from: /Users/xxxxx/work/test_dynamic_linking/./example
  Expected in: flat namespace
 in /Users/xxxxx/work/test_dynamic_linking/./example
Abort trap: 6

What I'm trying to achieve is to turn off build time linking, but dynamic loading libpython in runtime with dlopen. Note the -Wl,-undefined,dynamic_lookup flag in the cmake file. I'm doing it this way because I want to build a wheel, and linking to libpython is not a good idea AFAIU.

Below is a minimial reproducible example. I'm confused that if you call functions like Py_DecodeLocale() or Py_InitializeEx() directly from main.cpp, it works fine. But calling pybind11::initialize_interpreter() fails with the error above. If I do

nm -gU /opt/anaconda3/envs/py38/lib/libpython3.8.dylib | grep PyBaseObject

the symbol _PyBaseObject_Type is indeed defined in the lib:

0000000000334528 D _PyBaseObject_Type

If I create a wrapper shared library which wraps the calls to pybind11 functions, and dlopen it from main.cpp, it works fine. This makes me more confused.

Cmake file:

cmake_minimum_required(VERSION 3.4)
project(example)
set (CMAKE_CXX_STANDARD 11)
set(UNDEFINED_SYMBOLS_IGNORE_FLAG "-Wl,-undefined,dynamic_lookup")
string(APPEND CMAKE_EXE_LINKER_FLAGS " ${UNDEFINED_SYMBOLS_IGNORE_FLAG}")
string(APPEND CMAKE_SHARED_LINKER_FLAGS " ${UNDEFINED_SYMBOLS_IGNORE_FLAG}")
include_directories(pybind11/include)
include_directories(/opt/anaconda3/envs/py38/include/python3.8)
add_library(pywrapper SHARED ${CMAKE_CURRENT_SOURCE_DIR}/wrapper.cpp)
add_executable(example main.cpp)

pyembed.hpp:

#pragma once

#include "pybind11/embed.h"

namespace py = pybind11;

void initialize_interpreter_func();


struct pybind_wrap_api {
    decltype(&initialize_interpreter_func) initialize_interpreter;
};

wrapper.cpp:

#include "pyembed.hpp"

#include <set>
#include <vector>
#include <iostream>

#include "pybind11/embed.h"
#include "pybind11/stl.h"

namespace py = pybind11;


void initialize_interpreter_func() {
    pybind11::initialize_interpreter();
}

pybind_wrap_api init_pybind_wrap_api() noexcept {
    return {
        &initialize_interpreter_func,
    };
}

__attribute__((visibility("default"))) pybind_wrap_api pybind_wrapper_api =
    init_pybind_wrap_api();

main.cpp:

#include <pybind11/embed.h> // everything needed for embedding                                                                                                    

#include "pyembed.hpp"

#include <stdlib.h>
#include <dlfcn.h>
#include <iostream>
#include <string>

namespace py = pybind11;
static void* pylib_handle = nullptr;
static void* pybind_wrapper_handle = nullptr;
pybind_wrap_api* wrappers = nullptr;

int main() {
    std::string path_libpython = "/opt/anaconda3/envs/py38/lib/libpython3.8.dylib";
    pylib_handle = dlopen(path_libpython.c_str(), RTLD_NOW | RTLD_GLOBAL);
    if(!pylib_handle) {
        std::cout << "load libpython failed..." << std::endl;
    } else {
        std::cout << "load libpython succeeded..." << std::endl;
    }

    std::string path_wrapper = "./libpywrapper.dylib";
    pybind_wrapper_handle = dlopen(path_wrapper.c_str(), RTLD_NOW | RTLD_GLOBAL);
    wrappers = static_cast<pybind_wrap_api*>(dlsym(pybind_wrapper_handle, "pybind_wrapper_api"));

    std::string pythonhome = "/opt/anaconda3/envs/py38";
    setenv("PYTHONHOME", pythonhome.c_str(), 1);
    std::string pythonpath = "/opt/anaconda3/envs/py38/lib/python3.8/site-packages";
    setenv("PYTHONPATH", pythonpath.c_str(), true);

    // this line will cause it to fail with the symbol not found error
    py::initialize_interpreter();
    // if comment out the previous line and do the following line, it works fine. I'm confused why is so. 
    //wrappers->initialize_interpreter();

    return 0; 
} 

Then do

cmake . && make && ./example