Modernize test-suite in sparse tensors

[aocsa]

Stack from ghstack:

• Eliminate global usage of torch.set_default_dtype in test_autograd #56446 Eliminate global usage of torch.set_default_dtype in test_autograd
• Eliminate global usage of torch.set_default_dtype in sparse test #56393 Eliminate global usage of torch.set_default_dtype in sparse test
• Modernize test-suite in sparse tensors  #56392 Modernize test-suite in sparse tensors

Fixes for gh-56371 and gh-56369

Differential Revision: D27913212

Note:
The following requests about testing listed here

#50937 (comment) - see modern COO testing
#50937 (comment) - avoid using numpy
#50937 (comment) - inefficient CSR samples

are being solved in this PR

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit 2cd9c73 (more details on the Dr. CI page):

---

💚 💚 Looks good so far! There are no failures yet. 💚 💚

---

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

[mruberry]

Hey @aocsa! Overall this looks pretty good; I have a couple suggestions and a question, but nothing major

[pearu]

Thanks @aocsa , this looks good.

I have several nits but most importantly, I suggest making test_sparse_csr_constructor method more thorough, in particular, ensure that all supported dtypes are covered as well as checking that the constructed CSR tensor have the expected dtype, device, and indices/values.

We may also want to extend these tests to other inputs such as lists and objects implementing array interface but atm this blocked by #56687 and #54187 .

[aocsa]

Thanks @aocsa , this looks good.

I have several nits but most importantly, I suggest making test_sparse_csr_constructor method more thorough, in particular, ensure that all supported dtypes are covered as well as checking that the constructed CSR tensor have the expected dtype, device, and indices/values.

We may also want to extend these tests to other inputs such as lists and objects implementing array interface but atm this blocked by #56687 and #54187 .

Thanks for the feedback @pearu, I addressed the comments and rebased this PR on top of the latest changes, polished a few things, like better CSR sample input generator see #56392 (comment) and added some tests, specifically related to factory tests. I think it is ready for final review and merging, let me know what you think. cc @mruberry

[pearu]

Thanks @aocsa ! LGTM, however, there is a bug in random_sparse_csr. Once this is fixed, I'll approve.

[pearu]

Thanks, @aocsa for the fixes.

I'll approve this PR despite the wrong crow_indices[-1] <= nnz check, see #56392 (comment) .
We'll need to resolve this incorrectness in the context of #56856 because the corresponding fix likely requires changes to the CSR construction algorithm and hence is OT for this PR.

@mruberry I think this PR is ready for merge (assuming that CI is fine with it). The PR has a high priority because other CSR-related PRs rely on this testing framework.

[facebook-github-bot]

@mruberry merged this pull request in 18c89a9.