How to allocate an OrtValue array directly on GPU (to emulate low-memory conditions)

[vadimkantorov]

I assume something like this would work: onnxruntime.OrtValue.ortvalue_from_numpy(np.zeros((40_000_000_000, ), 'uint8'), device_type='cuda', device_id=0)

But is it possible to bypass the np.zeros / np.empty and directly allocate a GPU tensor using ORT-only API (without resorting to torch / cupy / cuda_driver)?