[BUG] - segmentation fault occur when follow the tutorial

Add Link

https://pytorch.org/tutorials/recipes/intel_neural_compressor_for_pytorch.html

Describe the bug

Follow the tutorial, I write this code, and find that the segmentation fault occur when the tensor(and the GraphModule model) moved to cuda:0

# main.py
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np

# LeNet Model definition
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
        self.conv2 = nn.Conv2d(10, 20, kernel_size=5)
        self.conv2_drop = nn.Dropout2d()
        self.fc1 = nn.Linear(320, 50)
        self.fc1_drop = nn.Dropout()
        self.fc2 = nn.Linear(50, 10)

    def forward(self, x):
        x = F.relu(F.max_pool2d(self.conv1(x), 2))
        x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
        x = x.reshape(-1, 320)
        x = F.relu(self.fc1(x))
        x = self.fc1_drop(x)
        x = self.fc2(x)
        return F.log_softmax(x, dim=1)

def save_quant():
    model = Net()
    model.load_state_dict(torch.load('./lenet_mnist_model.pth'))
    print(model)
    model = model.to("cuda:0")
    model.eval()
    from torchvision import datasets, transforms
    test_loader = torch.utils.data.DataLoader(
        datasets.MNIST('./data', train=False, download=True,
            transform=transforms.Compose([ transforms.ToTensor(), ])
            ),
        batch_size=1)

    # launch code for Intel® Neural Compressor
    from neural_compressor.experimental import Quantization
    quantizer = Quantization("./conf.yaml")
    quantizer.model = model
    quantizer.calib_dataloader = test_loader
    quantizer.eval_dataloader = test_loader
    q_model = quantizer()
    # q_model = q_model.to("cuda:0")
    q_model.save('./output')

def good_quant():
    from neural_compressor.utils.pytorch import load
    model = Net()
    int8_model = load('./output', model)
    first = np.random.rand(1,1,28,28).astype(np.float32)
    first = torch.from_numpy(first)
    print(f"when tensor on={first.device}")
    x=int8_model(first)
    print(x)

def bad_quant():
    from neural_compressor.utils.pytorch import load
    model = Net()
    device = "cuda:0"
    int8_model = load('./output', model).to(device)
    first = np.random.rand(1,1,28,28).astype(np.float32)
    first = torch.from_numpy(first).to(device)
    print(f"when tensor on={first.device}")
    x=int8_model(first)
    print(x)


save_quant()
good_quant()
bad_quant()

and the yaml file:

# conf.yaml
version: 2.0

model:
    name: LeNet
    framework: pytorch_fx

evaluation:
    accuracy:
        metric:
            topk: 1

tuning:
  accuracy_criterion:
    relative: 0.01

Expected Result: the quantized model run smoothly and efficiently on GPU

Actual Result: segmentation fault

Segmentation fault (core dumped)

Describe your environment

• Platform(Linux)
• CUDA(11, 12)
• 2.0.1+cu117

cc @ezyang @gchanan @zou3519 @kadeng @frank-wei @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @Xia-Weiwen @leslie-fang-intel

Start by reproducing main.py with conf.yaml and the linked Intel Neural Compressor tutorial in the stated Linux, CUDA 11/12, and PyTorch 2.0.1 environment. Compare save_quant, good_quant, and bad_quant to isolate the CUDA failure, then verify the tutorial example completes without a segmentation fault.