How can I bind torch tensor in GPU as input and output?

[braindevices]

I follow the instruction described in #10217

But I get cuda failure:

CUDA failure 700: an illegal memory access was encountered

Can anyone point out what I did wrong? or what is the correct way to use torch tensor in GPU?

device = torch.device("cuda:0")
input = torch.randn((batch, 3, hw[0], hw[1])).contiguous()
for _ in range(2):
    y = model(input)
output_shape = y.shape

import onnxruntime

providers = [
    (
        "CUDAExecutionProvider",
        {"device_id": device.index}
    )
]
session = onnxruntime.InferenceSession(model_onnx.SerializeToString(), providers=providers)
binding = session.io_binding()
numpy_dtype = input.numpy().dtype
binding.bind_input(
    name=session.get_inputs()[0].name,
    device_type=device.type,
    device_id=device.index,
    element_type=numpy_dtype,
    shape=input.shape,
    buffer_ptr=input.data_ptr())
y2 = torch.zeros_like(y).contiguous()
binding.bind_output(
    name=session.get_outputs()[0].name,
    device_type=device.type,
    device_id=device.index,
    element_type=numpy_dtype,
    shape=output_shape,
    buffer_ptr=y2.data_ptr())

session.run_with_iobinding(binding)
torch.cuda.synchronize()

[braindevices]

ah, I forgot to put the tensor to .cuda()