Generalize seeding for measurements

[rauletorresc]

Context: Currently, only catalyst.measure(), qml.sample() and qml.counts() can be seeded with qjit(seed=N). It would be nice to be able to seed all circuit readout methods, like probs(), expval(), var(), when executing a device with shots.

Description of the Change: Let the base measurement class store the device seed so it can be propagated when necessary to newly created derived measurements. Move the logic from the simulator to the base measurement class.

Benefits: All measurements using the generate_samples method will be seeded.

Measures to test locally:

• expval
• var
• probs
• sample
• counts

Devices to modify

• LQ
• LK
• LG

TODO

• Simplify design
• Document changes
• Test probs using the LQ, LK and LG plugins
• Test expval using the LQ, LK and LG plugins
• Test var using the LQ, LK and LG plugins
• Update changelog

[sc-77124]

[github-actions]

Hello. You may have forgotten to update the changelog!
Please edit .github/CHANGELOG.md with:

• A one-to-two sentence description of the change. You may include a small working example for new features.
• A link back to this PR.
• Your name (or GitHub username) in the contributors section.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 91.65%. Comparing base (074ddcf) to head (b46a8de).
Report is 1 commits behind head on master.

❗ There is a different number of reports uploaded between BASE (074ddcf) and HEAD (b46a8de). Click for more details.

HEAD has 57 uploads less than BASE

Flag	BASE (074ddcf)	HEAD (b46a8de)
	65	8

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1003      +/-   ##
==========================================
- Coverage   97.71%   91.65%   -6.06%     
==========================================
  Files         228      176      -52     
  Lines       36479    24744   -11735     
==========================================
- Hits        35645    22680   -12965     
- Misses        834     2064    +1230     

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

[dime10]

This looks like a clean solution to me! 💯 Have you tested that the randomness is now completely fixed for all measurements (e.g. decimals do not vary in an expval etc)?

[rauletorresc]

This looks like a clean solution to me! 💯 Have you tested that the randomness is now completely fixed for all measurements (e.g. decimals do not vary in an expval etc)?

Sorry to ask, but how can I test this :)

[paul0403]

This looks like a clean solution to me! 💯 Have you tested that the randomness is now completely fixed for all measurements (e.g. decimals do not vary in an expval etc)?

Sorry to ask, but how can I test this :)

https://github.com/PennyLaneAI/catalyst/blob/main/frontend/test/pytest/test_seeded_qjit.py

What I would do is on your local branch you can build catalyst first (make all ), this installs whatever lightning specified in wherever, then you can build your own lightning branch locally so it overwrites the previous installation

[dime10]

This looks like a clean solution to me! 💯 Have you tested that the randomness is now completely fixed for all measurements (e.g. decimals do not vary in an expval etc)?

Sorry to ask, but how can I test this :)

https://github.com/PennyLaneAI/catalyst/blob/main/frontend/test/pytest/test_seeded_qjit.py

What I would do is on your local branch you can build catalyst first (make all ), this installs whatever lightning specified in wherever, then you can build your own lightning branch locally so it overwrites the previous installation

In addition to running those tests for existing functionality, we should check that something like

@qml.qnode(qml.device("lightning.qubit", wires=1, shots=100))
def circuit():
    qml.RY(0.7, wires=0)
    return qml.expval(qml.PauliZ(0))

produces the same decimals with repeated runs if seeded.

[paul0403]

Acatully, I just remembered you can test this directly in lightning-catalyst, and no need to go through frontend https://github.com/PennyLaneAI/pennylane-lightning/pull/927/files#diff-934a38a9af759b117f260f29c5b2732bee701b7cd3b49a4bdf22da7f69973ad8

Edit: although adding expval/probs to the catalyst frontend tests mentioned above might also be necessary, just to make sure nothing's wrong during mlir.

[rauletorresc]

@dime10 @paul0403 after the last changes, expval does produce the same decimals with repeated runs if seeded. I'll perform similar checks with the rest of the measures. Keep you posted.

[rauletorresc]

@dime10 @paul0403 after the last changes, expval does produce the same decimals with repeated runs if seeded. I'll perform similar checks with the rest of the measures. Keep you posted.

Confirmed also with var, probs, sample and counts. And I learned state does not work with finite shots.

[paul0403]

Can we have tests similar to https://github.com/PennyLaneAI/pennylane-lightning/pull/927/files#diff-934a38a9af759b117f260f29c5b2732bee701b7cd3b49a4bdf22da7f69973ad8? So tests with the lightning-catalyst interface, not just the lightning base classes.

[maliasadi]

Nice work @rauletorresc! Happy to review the complete version of the support

[paul0403]

Looks good! Thanks @rauletorresc for working on this! 💯

As a final sanity check, can you maybe build against the lightning in this branch, and do some catalyst frontend tests like https://github.com/PennyLaneAI/catalyst/blob/main/frontend/test/pytest/test_seeded_qjit.py?

For example,

    dev = qml.device("lightning.qubit", wires=2, shots=100)

    @qjit(seed=37)
    def workflow():
        @qml.qnode(dev)
        def circuit():
            qml.Hadamard(wires=[0])
            qml.RX(12.34, wires=[1])
            return qml.probs()

        return circuit(), circuit(), circuit(), circuit()

   trail1 = workflow()  # [probs10, probs11, probs12, probs13]
   trail2 = workflow()  # [probs20, probs21, probs22, probs23]

   for (probs1i, probs2i) in zip(trail1, trail2):
      assert probs1i == probs2i // or whatever the equivalent pytest syntax is

Note that within the same workflow() call the four circuit() calls should give different results, i.e. the four probs1* and the four probs2* should be different among themselves.

(This should also be a PR in catalyst, after this lightning PR is merged)

[rauletorresc]

Looks good! Thanks @rauletorresc for working on this! 💯

As a final sanity check, can you maybe build against the lightning in this branch, and do some catalyst frontend tests like https://github.com/PennyLaneAI/catalyst/blob/main/frontend/test/pytest/test_seeded_qjit.py?

For example,

    dev = qml.device("lightning.qubit", wires=2, shots=100)

    @qjit(seed=37)
    def workflow():
        @qml.qnode(dev)
        def circuit():
            qml.Hadamard(wires=[0])
            qml.RX(12.34, wires=[1])
            return qml.probs()

        return circuit(), circuit(), circuit(), circuit()

   trail1 = workflow()  # [probs10, probs11, probs12, probs13]
   trail2 = workflow()  # [probs20, probs21, probs22, probs23]

   for (probs1i, probs2i) in zip(trail1, trail2):
      assert probs1i == probs2i // or whatever the equivalent pytest syntax is

Note that within the same workflow() call the four circuit() calls should give different results, i.e. the four probs1* and the four probs2* should be different among themselves.

(This should also be a PR in catalyst, after this lightning PR is merged)

Something like that I am doing at the Catalyst side here: PennyLaneAI/catalyst#1344
It works locally, I hope it will work in the CI once this PR gets merged.

[paul0403]

💯 🥳

All good from my side! I'm unfamiliar with the lightning codecov so maybe someone from performance can suggest there.

[maliasadi]

Nice work @rauletorresc! Happy to approve 🥳