Migrating Pytorch backend to Qiboml

[MatteoRobbiati]

Checklist:

• Reviewers confirm new code works as expected.
• Tests are passing.
• Coverage does not decrease.
• Documentation is updated.

Some things to be discussed and still not touched in this PR

1. I think we should fix the procedure based on this function, otherwise we will always need to depend on Torch;
2. In examples like this one we define a torch tensor which is provided to our circuit. I would like to drop any dependence on pytorch in Qibo, thus I propose to delete these kind of advanced examples, pointing to Qiboml documentation (where we will rewrite dedicated examples). Otherwise, we should build the torch tensor doing something like the following (which is very confusing and we must avoid):

backend = construct_backend(backend="qiboml", platform="pytorch")

# create a torch tensor from backend using torch which is renamed np
tensor = backend.np.ones( .... )

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.69%. Comparing base (ea6984d) to head (832d43c).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1510      +/-   ##
==========================================
- Coverage   99.72%   99.69%   -0.03%     
==========================================
  Files          80       79       -1     
  Lines       11784    11631     -153     
==========================================
- Hits        11752    11596     -156     
- Misses         32       35       +3     

Flag	Coverage Δ
unittests	99.69% <100.00%> (-0.03%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

---

🚨 Try these New Features:

• Flaky Tests Detection - Detect and resolve failed and flaky tests

[renatomello]

Should this be updated to contemplate the CPU / GPU subdivision? Both qibojit and qiboml have backends that work on both.

[MatteoRobbiati]

I don't know, since this is already shown in the backends diagram.

[renatomello]

1. I think we should fix the procedure based on this function, otherwise we will always need to depend on Torch;

That function is needed to properly have some methods of gates.Unitary working for all possible input types the matrix can be (i.e. np.ndarray, tf.Tensor, torch.Tensor, cp.ndarray). I'm not sure there is a way out of this dependency without creating a gate dependency on backends. One immediate thing we could do is to rename np as platform in all backends since that's the name we give in set_backend anyway.

[BrunoLiegiBastonLiegi]

Not sure whether it is worth to have details about the ml backends. I would probably place this in the qiboml docs and link them here.

[MatteoRobbiati]

I agree, I would remove these lines. Probably it is better to wait the moment the Qiboml docs are ready so that we can just link to the proper section.

[BrunoLiegiBastonLiegi]

This is not really for this PR, but I think that both of them support sparse matrices to some extent https://pytorch.org/docs/stable/generated/torch.sparse.mm.html
https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/sparse-mat-mul

I am not sure what was causing problems here, but I guess it is more a matter of how the operation is called rather than the impossibility to perform it. Anyway, this is probably out of the scope of this PR.

[MatteoRobbiati]

I just opened an issue with this comment.

[BrunoLiegiBastonLiegi]

This could become a property of the backend, e.g.:

@property
def support_differentiation(self,):
    return False

to avoid checking precisely for which backend this works, but again might be out of scope here.

[MatteoRobbiati]

I think the best thing for Qibo will be to remove all these tests about variational models.

[MatteoRobbiati]

By my side, I think we can stop modifications here for this PR. Still we need to remove a lot of examples and subsections from the docs, but I would prefer to do that after we populate the Qiboml documentation. Otherwise we will have vacations in the documentation. What you think @renatomello, @BrunoLiegiBastonLiegi?

[BrunoLiegiBastonLiegi]

We need to update this https://github.com/qiboteam/qiboml/blob/800e800d54f4962f4890716c4b1539014c69c1b5/src/qiboml/backends/__init__.py#L3

BTW, is the PytorchBackend of qiboml updated or we are still waiting for qiboteam/qiboml#30 ?

[MatteoRobbiati]

We need to update this https://github.com/qiboteam/qiboml/blob/800e800d54f4962f4890716c4b1539014c69c1b5/src/qiboml/backends/__init__.py#L3

Done in qiboml, thanks!

BTW, is the PytorchBackend of qiboml updated or we are still waiting for qiboteam/qiboml#30 ?

That was outdated. Just closed!

[BrunoLiegiBastonLiegi]

okay it should be just a matter of regenerating the lock here now

[renatomello]

qibo is importing the torch backend from qiboml and qiboml is importing the torch backend from qibo

https://github.com/qiboteam/qibo/actions/runs/11888475398/job/33209011064?pr=1510

[MatteoRobbiati]

qibo is importing the torch backend from qiboml and qiboml is importing the torch backend from qibo

This should have been solved merging qiboteam/qiboml#52 (comment) yesterday. Rerunning the tests to double check :)

[MatteoRobbiati]

I just removed a gradient test from here and opened a dedicated issue in qiboml.
Also, the test_list_available_backends should be fixed here, so after merging the master into your branch, @renatomello, the problem mentioned in #1520 should be solved.

[renatomello]

I just removed a gradient test from here and opened a dedicated issue in qiboml. Also, the test_list_available_backends should be fixed here, so after merging the master into your branch, @renatomello, the problem mentioned in #1520 should be solved.

You mean after this PR is merged, right?

[MatteoRobbiati]

You mean after this PR is merged, right?

Yes, I mean this PR should be solving that problem too.