VQE Runtime ignores any parameters of optimizer other than SPSA/QNSPSA

[LenaPer]

Describe the bug
When running a vqe runtime through QiskitRuntimeService(channel='ibm_quantum') with an optimizer other than SPSA/QNSPSA, and looking at the logs (job.logs()), I noticed the optimizer is set to be SLSQP, even when I pass any other optimizer, it shows on the logs :
2022-07-04T09:37:35.046460730Z -- optimizer: <qiskit.algorithms.optimizers.slsqp.SLSQP object at 0x7fed33b72d60>
And later in the logs :

2022-07-04T09:37:35.046639714Z ===============================================================
2022-07-04T09:37:35.046645335Z Optimizer: SLSQP
2022-07-04T09:37:35.046650890Z -- method: slsqp
2022-07-04T09:37:35.046656375Z -- bounds_support_level: 2
2022-07-04T09:37:35.046661230Z -- gradient_support_level: 2
2022-07-04T09:37:35.046665764Z -- initial_point_support_level: 3
2022-07-04T09:37:35.046672463Z -- options: {'maxiter': 100, 'disp': False, 'ftol': 1e-06, 'eps': 1.4901161193847656e-08}
2022-07-04T09:37:35.046677546Z -- max_evals_grouped: 1
2022-07-04T09:37:35.046683473Z -- kwargs: {'tol': None}
2022-07-04T09:37:35.046689198Z ===============================================================

One of the job ids about this : cb1fncea4p74378e70v0

Steps to reproduce
The code to reproduce the error :

import qiskit
from qiskit.opflow import Z
from qiskit.circuit.library import TwoLocal
from qiskit_ibm_runtime import QiskitRuntimeService
from qiskit.algorithms.optimizers import COBYLA

service = QiskitRuntimeService(channel='ibm_quantum')

n = 2
operator = Z^Z
ansatz = TwoLocal(n, 'ry', 'cx', 'linear', reps=1)

runtime_inputs_cobyla = {
    'ansatz': ansatz,
    'operator': operator, 
    'optimizer': COBYLA(maxiter=1), 
}

options = {
    'backend_name': 'ibmq_qasm_simulator'
}

job_cobyla = service.run(
    program_id='vqe',
    options=options,
    inputs=runtime_inputs_cobyla,
)

Expected behavior
The right optimizer shows in the logs

Suggested solutions

Additional Information

• qiskit-ibm-runtime version: 0.5.0
• Python version: 3.9.12
• qiskit version: 0.37

[rathishcholarajan]

@Cryoris Could you please help check?

[LenaPer]

Weirdly enough without changing anything on my environment I notice now the optimiser is not ignored anymore, but I'm still seeing that when using any optimizer other than SPSA/QNSPSA, any parameters I pass to the optimizer gets ignored and it uses the default ones (looking at this via job.logs())

[Cryoris]

Yeah we updated the VQE program and alongside fixed that the optimizers are ignored 🙂 Ok the lost settings are certainly a bug though, not quite sure where this originates from since the optimizers have a settings attribute which the runtime JSON encoder and decoders should use to reconstruct the object...

[LenaPer]

As an example, when I pass COBYLA(maxiter=1), it shows in the logs it still has the default parameters

2022-07-08T14:06:50.003629122Z Optimizer: COBYLA
2022-07-08T14:06:50.003629122Z -- method: cobyla
2022-07-08T14:06:50.003629122Z -- bounds_support_level: 1
2022-07-08T14:06:50.003629122Z -- gradient_support_level: 1
2022-07-08T14:06:50.003629122Z -- initial_point_support_level: 3
2022-07-08T14:06:50.003629122Z -- options: {'disp': False, 'maxiter': 1000, 'rhobeg': 1.0}
2022-07-08T14:06:50.003631508Z -- max_evals_grouped: 1
2022-07-08T14:06:50.003636219Z -- kwargs: {'tol': None}

[LenaPer]

Any updates on this?

[Cryoris]

Not yet -- I'll try having a look this week 😄

[Cryoris]

This issue seems to come from the encoding and decoding the optimizer which reset the settings. For example, the following serialization and deserialization of COBYLA

import json
from qiskit.algorithms.optimizers import COBYLA
from qiskit.providers.ibmq.runtime import RuntimeEncoder, RuntimeDecoder

obj = COBYLA(maxiter=1)
print("Original:\n", obj.settings)
dumped = json.dumps(obj, cls=RuntimeEncoder)
loaded = json.loads(dumped, cls=RuntimeDecoder)
print("Loaded:\n", loaded.settings)

prints

Original:
 {'max_evals_grouped': 1, 'options': {'maxiter': 1, 'disp': False, 'rhobeg': 1.0}, 'tol': None}
Loaded:
 {'max_evals_grouped': 1, 'options': {'maxiter': 1000, 'disp': False, 'rhobeg': 1.0}, 'tol': None}

It seems this comes from the problem that SciPy-based optimizers cannot be reconstructed from their own settings

>>> COBYLA(**obj.settings).settings
{'max_evals_grouped': 1, 'options': {'maxiter': 1000, 'disp': False, 'rhobeg': 1.0}, 'tol': None}

which is an assumption we have with the runtime encoder and decoders.

[Cryoris]

Ok something really funky is going on (keep track of maxiter)

>>> from qiskit.algorithms.optimizers import COBYLA
>>> one_iter_optimizer = COBYLA(maxiter=1)
>>> one_iter_optimizer.settings
{'max_evals_grouped': 1, 'options': {'maxiter': 1, 'disp': False, 'rhobeg': 1.0}, 'tol': None}
>>> COBYLA(**one_iter_optimizer.settings)
<qiskit.algorithms.optimizers.cobyla.COBYLA object at 0x7f82100709d0>
>>> one_iter_optimizer.settings
{'max_evals_grouped': 1, 'options': {'maxiter': 1000, 'disp': False, 'rhobeg': 1.0}, 'tol': None}

[Cryoris]

Qiskit/qiskit/pull/8343 (plus backport and update of the runtime images) should fix this 🙂

[kt474]

This issue has been resolved