Can't call tensor.data_ptr() with torch_directml on DML device

[SmartWashingMachine]

When a tensor is moved / initially placed in the DML device it seems the data pointer can't be read. Instead it throws the error:
RuntimeError: Cannot access data pointer of Tensor that doesn't have storage

Is there a way around this? I'm hoping to use the data pointer to pre-allocate inputs and outputs and use ONNX Runtime's IO Binding API with DirectML.

Basic code snippet to reproduce:

import torch
import torch_directml
dml = torch_directml.device(0)

x = torch.tensor([1, 2]).to(dml)
print(x.data_ptr())

[zhangxiang1993]

Hi,

Unfortunately, DML Tensor doesn't have storage by design because DmlTensorImpl inherits the OpaqueTensorImpl which assumes that storage is never set. And the reason behind that is we don't access DML memory by indexing which is extremely slow in DirectML.