[PyTorch Deprecation] Tensor.storage() has been deprecated

[yaox12]

Description

Tensor.storage() has been deprecated in PyTorch master and this change would be shipped with PyT 2.0. We may need to update the code calling Tensor.storage() in DGL, because it will raise a warning and finally cause an error.

Note that Tensor._storage() has been renamed to Tensor.untyped_storage() in pytorch/pytorch#91414, which makes it difficult to use a consistent API among PyTorch versions DGL's supporting.

To Do

• zerocopy_to_dgl_ndarray_for_write fixed by [Fix] Fix tensor.storage() deprecation warning #5656.
• InvertibleCheckpoint

[jermainewang]

@peizhou001 Could you take a look? We should first locate where in DGL are we using Tensor.storage().

[github-actions]

This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you

[yaox12]

DGL is using Tensor.storage() in two places:

1. in zerocopy_to_dgl_ndarray_for_write, which is easy to fix

   dgl/python/dgl/backend/pytorch/tensor.py

   Line 449 in 9cb6693

   assert input.numel() == input.storage().size(), (

2. in InvertibleCheckpoint, which I'm not familiar with. Can you take a look to see if there's a workaround? @mufeili

   dgl/python/dgl/nn/pytorch/conv/grouprevres.py

   Line 35 in 9cb6693

   inputs[1].storage().resize_(0)

[Rhett-Ying]

close as fixed