Pearu Peterson

Should the rename happen in-place or via introducing a new feedstock and the old will be archived? https://github.com/conda-forge/arrow-cpp-feedstock/pull/158#issuecomment-650537999

IIUC, it would be wrong to merge this PR until comments in https://github.com/pytorch/pytorch/pull/85378#pullrequestreview-1119751425 are addressed and the #85378 is approved.

To answer some of the "why" questions here: In general, operators on sparse tensors return a coaleased or uncoalesced sparse tensor. Which one is returned, depends on the operator and...

@vadimkantorov If you don't mind creating an issue for torch.save/load vs sparse coo tensor discrepancy, then that would be great!

With the current master we have: ```python >>> import torch >>> a = torch.zeros((12, 1, 1), layout=torch.sparse_coo) >>> a tensor(indices=tensor([], size=(3, 0)), values=tensor([], size=(0,)), size=(12, 1, 1), nnz=0, layout=torch.sparse_coo) >>>...

`nnz` in matlab is totally fine as it is defined as `(tensor != 0).sum()`. For non-hybrid sparse tensors, that would be `(tensor.values() != 0).sum()`. The problem is in sparse formats...

> e.g. when saving a narrow'd big tensor, torch.save still saves the full tensor without cloning - i would argue that this is even more confusing I agree. > than...

As of date, reading and writing binary data is supported.

Thanks for the PR! Can you address the CI failures? Sorry for the late response.

Is this still the issue?