Update _safe_divide to allow Accuracy to run on the GPU

[ndrwrbgs]

Caused a CPU/GPU sync point. torch.where can accept the float directly

What does this PR do?

Fixes #2638

Before submitting

• Was this discussed/agreed via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

Ensure CPU/GPU sync point avoided for Accuracy metric

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

Make sure you had fun coding 🙃

---

📚 Documentation preview 📚: https://torchmetrics--2640.org.readthedocs.build/en/2640/

[ndrwrbgs]

Oh! Thank you @Borda for doing cleanups for me. I apologize I didn't get pre-commit installed until after the initial commit.

[Borda]

Oh! Thank you @Borda for doing cleanups for me. I apologize I didn't get pre-commit installed until after the initial commit.

no worries :)

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 69%. Comparing base (bd777c9) to head (721bcf8).
Report is 103 commits behind head on master.

Additional details and impacted files

@@          Coverage Diff           @@
##           master   #2640   +/-   ##
======================================
  Coverage      69%     69%           
======================================
  Files         316     316           
  Lines       17881   17881           
======================================
  Hits        12331   12331           
  Misses       5550    5550           

[Borda]

@ndrwrbgs mind checking the failing tests?

[ndrwrbgs]

Of course! I’m grateful for the good test coverage and can see what’s happening here.

Downgrading to draft so it’s off the docket until it is ready, I overestimated the ease.

[Borda]

Downgrading to draft so it’s off the docket until it is ready, I overestimated the ease.

as a draft you will not run all tests so as it is at its advanced stage, let's keep it as full PR just add wip in the title

[SkafteNicki]

@ndrwrbgs the error only happens for older versions of PyTorch. The problem is that torch.where:

torch.where(condition, input, other, *, out=None)

expects other and input to have same dtype in older versions of Pytorch whereas in newer versions of Pytorch it will automatically upcast to the highest precision. So we most likely need to explicit cast zero_division to the same dtype as num/denom

[ndrwrbgs]

Thank you! Promise I’ll get it fixed, I opened this just before visiting family and have gotten busy 😊.

[ndrwrbgs]

Thank you for your patience!

---

I noticed errors like below in this Azure job:

error:  cannot create tests/_cache-references/tests/_cache-references/.unittests.retrieval.helpers._compute_sklearn_metric_276bbfaa62e3d293c69b71ebae3d3e4a266434eef12c60faa590f4111afbb6f7
        No space left on device

Here's an example both from my build and another before to confirm it's not introduced with this PR:

• build 1
• build 2

🤔 I'm not sure whether it has any observable impact or not.

cc @Borda

[ndrwrbgs]

cc @SkafteNicki Another failure on the Windows 2.4.0 torch.Unique: https://github.com/Lightning-AI/torchmetrics/actions/runs/10150851419/job/28068885801?pr=2640

@Borda the other failures appear to be unrelated to the change, but I'm not seeing any rerun button available on my end. Is that something you're able to to with write permissions?