Support V2 primitives

[woodsp-ibm]

What should we add?

The Sampler and Estimator are being updated with new V2 versions that change the interface and will be available from Qiskit 1.0.0. V1 versions will at some point be deprecated and removed. Algorithms should add support for V2 versions and ultimately drop support for V1 once they are removed.

[tnemoz]

Is it possible to submit some small PR to help supporting V2 primitives? Like, would a PR allowing VQE to support both V1 and V2 be useful? Or is it better to wait for qiskit-community/qiskit-machine-learning#817 to be merged? Or, another possibility, is it something you're already working on and the PR won't help much?

[woodsp-ibm]

@tnemoz There is no work being done by present maintainers (IBM) on V2 (or ISA #164) changes for algorithms. See disclaimer in readme

Qiskit Algorithms is no longer officially supported by IBM.

Like any other Apache 2 licensed code, you are free to use it or/and extend it, but please be aware that it is under your own risk.

The Qiskit Machine Learning maintainers, in that PR you linked, they have copied just the parts from algos that are needed/used in ML to be able to maintain/support them there going forwards so its independent. Qiskit Optimization has a PR that does likewise qiskit-community/qiskit-optimization#623

I do not think either will change VQE though as neither use it. There is a fair amount overall to be changed here, if algorithms were kept like it is and to be maintained going forwards.

[LenaPer]

Would it be possible to be assigned to this issue with @tnemoz ? We'd like to give it a try if possible.

[tnemoz]

The new BaseSamplerV2 and BaseEstimatorV2 from Qiskit don't have an option attribute like the V1 used to, while the BasePrimitiveV2 from qiskit-ibm-runtime does.

Is it worth it to add qiskit-ibm-runtime as a dependency and to check with isinstance whether it is actually possible to set these options? Something along the lines of:

if options is not None:
    if not isinstance(estimator, qiskit_ibm_runtime.BaseEstimatorV2):
        raise ValueError("options is only supported for a runtime Estimator")

This problem happens for instance for migrating ComputeUncompute to a V2 Sampler.

[woodsp-ibm]

@LenaPer @tnemoz I assigned the issue to you as requested. Now clearly to support runtime it needs the ISA change as well. Very shortly V1 primitives will be removed from runtime (they have been deprecated there for a while) and those remaining in Qiskit (the reference ones that run classically) will be deprecated. As was noted above Qiskit Machine Learning has copied over what was used from here, Qiskit Optimization likewise. ML has yet to do anything about ISA and V2 while Opt does have a draft PR addressing things. If you go ahead and look at things, perhaps in more depth than you might have done so far, and decide you still want to move forwards perhaps outline what/how you want to change things to have a discussion up front before really doing too much. In regards of the runtime comment above, in places we do have optional install checks so that the check for runtime could be done more conditionally based on it being installed or not, rather than having it as a hard dependency.

[tnemoz]

If you go ahead and look at things, perhaps in more depth than you might have done so far, and decide you still want to move forwards perhaps outline what/how you want to change things to have a discussion up front before really doing too much.

The goal would be to remove the support for V1 primitives since those are going to be deprecated and add ISA support, in a similar fashion to what's being done in the qiskit-optimization PR. Eventually, I think the goal would be to be able to use every algorithm in here with the Runtime and Aer primitives (as noted above, the qiskit ones have a somewhat different interface than these two) with the least friction possible to move from one to another.

If that's OK with you, we'll issue a Draft PR shortly that will address only a short part of the issue (probably on minimal eigensolvers or phase estimation), just so that we can discuss with actual examples whether things move in the right direction according to you. That would avoid a painful review where the same mistake is repeated throughout all files.