Fix deepcopy of ov.Tensor #1146
Conversation
optimum/intel/openvino/utils.py
Outdated
| elif isinstance(inputs, tuple): | ||
| new_inputs = tuple(deepcopy_data(elem) for elem in inputs) | ||
| elif isinstance(inputs, openvino.Tensor): | ||
| new_inputs = openvino.Tensor(np.zeros(inputs.shape, dtype=inputs.element_type.to_dtype())) |
There was a problem hiding this comment.
| new_inputs = openvino.Tensor(np.zeros(inputs.shape, dtype=inputs.element_type.to_dtype())) | |
| new_inputs = openvino.Tensor(np.empty(inputs.shape, dtype=inputs.element_type.to_dtype())) |
optimum/intel/openvino/utils.py
Outdated
| elif isinstance(inputs, tuple): | ||
| new_inputs = tuple(deepcopy_data(elem) for elem in inputs) | ||
| elif isinstance(inputs, openvino.Tensor): | ||
| new_inputs = openvino.Tensor(np.zeros(inputs.shape, dtype=inputs.element_type.to_dtype())) |
There was a problem hiding this comment.
What will happen if inputs.element_type is bf16, u4 or i4?
There was a problem hiding this comment.
Thanks for the comment!
I added a type argument to be sure that output tensor has a correct type.
tests/openvino/test_quantization.py
Outdated
| "f": [ov.Tensor(np.ones((1, 2, 3))), ov.Tensor(np.ones((1, 2, 3)))], | ||
| } | ||
| copied_data = deepcopy_data(data) | ||
| assert copied_data["a"] is not data["a"] |
There was a problem hiding this comment.
Please use self.assertTrue(...) or similar instead of assert ...
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
The right place for this fix is OpenVINO Python API. If we do it we can fix it in all the integrations of OV to 3rd party projects including Optimum, Ulralytics, etc. |
There is an open ticket to add |
I am still thinking that this is not a good idea to fix OV issues in 3rd party projects. From what I understood you have problems with NNCF CI. To address them quickly, you could:
cc'ed @alexsu52 to drive that |
What does this PR do?
Fix applying deepcopy to ov.Tensor