import Pytorch to create Neural Network. Running Error

I'm using C++/Qt to run my Qt application. Development Version: Python: 3.12 Qt: 6.6.3 C++: std 17 Visual Studio 2022

Inside my code, I use Python/C API to run python script like:

   PyObject* pModule = PyImport_ImportModule("DataModule");
   if (pModule == nullptr)
   {
       PyErr_Print();
       return QString();
   }

   PyObject* pFunc = PyObject_GetAttrString(pModule, "RunNeuralNetwork");

   PyObject* py_outputDir = QStringToPyObject(outputDir);
   PyObject* py_fileName = QStringToPyObject(fileName);


   PyObject* pArgs = PyTuple_New(4);
   PyTuple_SetItem(pArgs, 0, py_outputDir);
   PyTuple_SetItem(pArgs, 1, py_fileName);

   PyObject* pReturn = PyObject_CallObject(pFunc, pArgs);

   QString qString = PyObjectToQString(pReturn);

In "DataModule.py" file, I import Pytorch and create a Neural NetWork function. the code like:

import torch
from torch import nn
from torch.func import jacrev
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
from multiprocessing import Pool

...

class NeuralNetwork(nn.Module):

    def __init__(self,
                 dim_x: int,
                 dim_y: int,
                 num_hidden_layer: int,
                 num_node: int,
                 acti_name: str,
                 kwarg_acti: dict = None,
                 use_bn: bool = False,
                 normal=None) -> None:
        super().__init__()
        activation_func = getattr(nn, acti_name)
        if kwarg_acti is None:
            kwarg_acti = dict()

        dim_list = [dim_x] + [num_node] * num_hidden_layer
        layer_list = []
        for i in range(num_hidden_layer):
            layer_list.append(nn.Linear(dim_list[i], dim_list[i + 1]))
            if use_bn:
                layer_list.append(nn.BatchNorm1d(dim_list[i + 1]))
            layer_list.append(activation_func(**kwarg_acti))
        layer_list.append(nn.Linear(dim_list[i + 1], dim_y))
        # layer_list.append(nn.BatchNorm1d(dim_y))
        self.layers = nn.Sequential(*layer_list)

        if normal is None:
            self.loc_x = torch.tensor(0.)
            self.loc_y = torch.tensor(0.)
            self.sca_x = torch.tensor(1.)
            self.sca_y = torch.tensor(1.)
        else:
            self.loc_x, self.loc_y, self.sca_x, self.sca_y = normal

    def forward(self, x):
        x = self.layers((x - self.loc_x) / self.sca_x)
        return x * self.sca_y + self.loc_y

When I run this code in C++ Python/C API, sometimes it runs well, and sometimes it crash.

It can not work neither in main thread or child thread.

The crash info in VS is:

0x00007FFE06A5EE30 (python311.dll) Exception in MyApp.exe 
0xC0000005: access violation reading 0xFFFFFFFFFFFFFFFF

I consider it is the problem of Multi Thread, then I use the below code In '.py' script

torch.set_num_threads(1)

It still not work.

Please Help!

use C++ through Python/C API runs '.py' script importing pytorch and no error.