QPY serialization of ParameterExpression throws error

[jlapeyre]

• Qiskit version: 1.3.2
• qiskit-ibm-runtime version: 0.36.1
• Python version: 3.12

Certain ParameterExpressions cause QPY serialization v13 to throw. QPY v12 does not throw.

Note

After opening this issue I found the following much simpler script for triggering the bug:

from qiskit.circuit import Parameter, ParameterExpression
from qiskit import qpy

a = Parameter("a")
b = Parameter("b")
a1 = a * 2
a2 = a1.subs({a: 2 * b})

qpy.binary_io.value.dumps_value(a2, version=13);

There is no error with QPY version 12.

When running the script above, the error,

AttributeError: 'ParameterExpression' object has no attribute 'name'

is thrown here at line 40

qiskit/qiskit/qpy/binary_io/value.py

Lines 39 to 42 in d9b8a18

	def _write_parameter(file_obj, obj):
	name_bytes = obj.name.encode(common.ENCODE)
	file_obj.write(struct.pack(formats.PARAMETER_PACK, len(name_bytes), obj.uuid.bytes))
	file_obj.write(name_bytes)

I tried to track down the problem, which is centered on _qpy_replay. But there are no documentation or comments describing _qpy_replay and several adjacent structures and functions.

Note

The following script is not a minimal description of the bug. But it is an example abstracted from using EstimatorQNN and NeuralNetworkClassifier from qiskit-machine-learning on a real IBM device (in this case, ibm_kyiv). It's unlikely that simple workflows involving EstimatorQNN will work when this bug is present.

Running this script will trigger the bug:

from qiskit.circuit.library import ZZFeatureMap, RealAmplitudes
from qiskit import QuantumCircuit
from qiskit_ibm_runtime import fake_provider
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager

def make_circuit():
    n_qubits = 2
    reps = 1
    feature_map = ZZFeatureMap(feature_dimension=n_qubits, reps=reps)
    ansatz = RealAmplitudes(num_qubits=n_qubits, reps=reps)
    qc = QuantumCircuit(n_qubits)
    qc.compose(feature_map, inplace=True)
    qc.compose(ansatz, inplace=True)
    backend = fake_provider.FakeKyiv()
    pm = generate_preset_pass_manager(backend=backend, optimization_level=1)
    isa_qc = pm.run(qc)
    return isa_qc

from qiskit import qpy
isa_qc = make_circuit()
qpy.binary_io.value.dumps_value(isa_qc.global_phase, version=13);

If I replace versions=13 with version=12, I see no apparent problem.

[mtreinish]

I checked and this bug is present in the public api for qpy too. Calling qiskit.qpy.dump() will trigger the bug if the parameter expression as defined in the reproducer is used on a gate. qpy.binary_io.value.dumps_value is not part of the public api so I just wanted to rule out that it's not an issue from calling the internals directly.

[jlapeyre]

@mtreinish, that's important. It would have been better if I included an explicit user-level API example in the bug report. I think both things are useful. Showing that a user will encounter the bug, and a minimal example.

This came up in code from collaborators. It's a simple machine learning example. I think most of the code is from a tutorial. It works fine on a fake backend. But fails with a real backend because of the qpy serialiazation. On the application side, QNNEstimator has only been supported on real V2 backends for a couple of months. I have not seen any other mention of people running into this problem.