Segfault when using IO binding to CUDA tensor with CPU execution provider

Describe the issue

When running an inference session with a binding to GPU memory and using the CPU execution provider, onnxruntime will segfault. I would expect a helpful error message instead.

To reproduce

import onnxruntime as ort
import torch
import torch.nn as nn
import numpy as np

class MyModel(nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.fc1 = nn.Linear(4, 2)

    def forward(self, x):
        return self.fc1(x)

torch_model = MyModel()
torch_input = torch.randn(1, 4)
onnx_model = torch.onnx.dynamo_export(torch_model, torch_input)

onnx_model.save("model.onnx")

sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
binding = sess.io_binding()

for n in sess.get_inputs():
    binding.bind_ortvalue_input(
            n.name,
            ort.OrtValue.ortvalue_from_numpy(
                np.zeros(shape=n.shape, dtype=np.float32)))

for n in sess.get_outputs():
    binding.bind_ortvalue_output(
            n.name,
            ort.OrtValue.ortvalue_from_numpy(
                np.zeros(shape=n.shape, dtype=np.float32),
                device_type='cuda'))

sess.run_with_iobinding(binding)

This exits with code 139 and prints:

Segmentation fault (core dumped)

If device_type is changed to cpu then it runs correctly.

Urgency

Not urgent.

Platform

Linux

OS Version

Fedora 40

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.19.0

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

microsoft / onnxruntime

Segfault when using IO binding to CUDA tensor with CPU execution provider #21865