Devices and backend boundaries

PennyLane 0.45.1 is the tested core. Provider integrations below were checked against official docs/released source, not authenticated hardware jobs. Install each provider/compiler in a separate environment and resolve its declared requirements; a plugin version number does not guarantee all current backend APIs.

Local simulators and reusable circuits

import pennylane as qml
from pennylane import numpy as np

def quantum_function(theta):
    qml.RY(theta, wires=0)
    qml.CNOT(wires=[0, 1])
    return qml.expval(qml.Z(1))

results = {}
for name in ["default.qubit", "lightning.qubit", "default.mixed"]:
    dev = qml.device(name, wires=2)
    circuit = qml.QNode(quantum_function, dev, diff_method="parameter-shift")
    theta = np.array(0.3, requires_grad=True)
    results[name] = float(circuit(theta))
    assert np.allclose(results[name], np.cos(theta))
    assert np.allclose(qml.grad(circuit)(theta), -np.sin(theta))

default.mixed supports explicit channels; changing to it does not automatically apply a realistic hardware noise model. Lightning speedups depend on circuit size and overhead. lightning.gpu requires its separate CUDA/cuQuantum-compatible plugin and hardware; no GPU fallback should silently change a reported benchmark. default.clifford needs Stim and accepts only its supported Clifford operations and measurements (some Pauli rotations at Clifford angles may be supported). There is no universal fixed qubit limit or provider noise ranking.

Use dev.name, dev.wires and len(dev.wires) when wires are allocated. New Device API objects do not universally expose legacy num_wires, operations, observables or apply. Some use dynamic wires or optional capabilities objects. Inspect each device's documented preprocessing and derivative support; run a tiny representative circuit. For custom devices implement the current execution and preprocessing protocol, not an obsolete DefaultQubit.apply override. A custom operator can implement compute_decomposition; registering an arbitrary attribute on qml.ops is not plugin registration.

Provider installation targets

These released versions were source/metadata checked on 2026-10-01. Commands are illustrative dependency setup, not a tested combined environment:

# Choose one provider per environment.
uv pip install "pennylane==0.45.1" "pennylane-qiskit==0.45.0"
uv pip install "pennylane==0.45.1" "amazon-braket-pennylane-plugin==1.35.2"
uv pip install "pennylane==0.45.1" "pennylane-cirq==0.44.0"
uv pip install "pennylane==0.45.1" "pennylane-ionq==0.45.0"

Rigetti's last published plugin is 0.40.0, with pyquil<4 and networkx<3; resolve and validate it in its own legacy stack instead of installing it into the current core/ML environment. Current QCS hardware availability requires provider verification. Strawberry Fields' plugin documentation targets 0.29.1; an old strawberryfields.remote/Borealis recipe is not evidence of compatibility or a currently accessible backend with PennyLane 0.45.1.

IBM Quantum: backend object and finite shots

qiskit.aer is a local simulator plugin. qiskit.remote accepts a Qiskit backend object and uses EstimatorV2/SamplerV2. Allocate only the logical wires needed, within backend capacity. Its execution is sampled; do not promise analytic results.

Illustrative authenticated setup, not executed:

import os
import pennylane as qml
from qiskit_ibm_runtime import QiskitRuntimeService

service = QiskitRuntimeService(
    channel="ibm_quantum_platform",
    token=os.environ["IBM_QUANTUM_API_KEY"],
    instance=os.environ["IBM_QUANTUM_INSTANCE"],
)
backend = service.least_busy(operational=True, simulator=False, min_num_qubits=2)
dev = qml.device("qiskit.remote", wires=2, backend=backend)
@qml.set_shots(1024)
@qml.qnode(dev, diff_method="parameter-shift")
def hardware_circuit(theta):
    qml.RY(theta, wires=0)
    qml.CNOT(wires=[0, 1])
    return qml.expval(qml.Z(1))
# Calling hardware_circuit submits work; record backend, calibration, shots and job metadata.

The environment variable names here are application-defined inputs, not automatic SDK credential discovery. Saved credentials are another option: QiskitRuntimeService.save_account(...) saves configuration; then construct QiskitRuntimeService() separately. Do not assign save_account's return value as the service. The SDK resolves the IBM platform URL and handles authentication; this skill does not implement a parallel REST client. Source docs and released plugin source differ on the behavior of unspecified shots, so set an explicit finite integer. Shot vectors are not supported by this plugin release.

Amazon Braket

The plugin provides braket.local.qubit for local execution and braket.aws.qubit for cloud tasks. Cloud construction uses device_arn, wires and optional s3_destination_folder=(bucket, prefix); use a bucket the account owns/accesses. AWS SDK credentials and correct region are required for cloud use. Discover the currently available device ARN with the Braket console/SDK and check status, supported operations, shots and result types before selecting it. Do not copy a historical Harmony ARN or an example bucket name into a live request. The documented SV1 ARN is arn:aws:braket:::device/quantum-simulator/amazon/sv1; a cloud simulator can still incur cost. No Braket task was submitted here.

Cirq, Rigetti and photonics

The Cirq 0.44.0 plugin registers cirq.simulator, cirq.mixedsimulator, cirq.qsim, cirq.qsimh and cirq.pasqal. qsim requires its optional native package. **cirq.pasqal describes Pasqal neutral-atom devices, not Google's Rainbow hardware.** Google Quantum Engine access needs an appropriate current provider integration and access; do not infer it from installing the Cirq plugin.

Rigetti's legacy entry points include rigetti.qvm, rigetti.qpu and wavefunction simulators. A QVM needs its runtime/services; a QPU needs current QCS credentials and a currently accessible processor. Do not hardcode retired Aspen targets or assume a QPU's number of qubits from an old example. Photonic workflows use different state spaces/operations; confirm the dedicated plugin/runtime instead of swapping a qubit QNode's device string to Borealis.

IonQ: explicit current backend target

The released pennylane-ionq==0.45.0 package registers ionq.simulator and ionq.qpu. It reads an explicit api_key, then PENNYLANE_IONQ_API_KEY, then IONQ_API_KEY. Its client uses IonQ API v0.4 and Authorization: apiKey .... An ionq.simulator circuit still uses the remote service and credentials.

For hardware, provide a currently discovered target explicitly. The released plugin's generic target="qpu" fallback still rewrites to qpu.aria-1, while current v0.4 create-job documentation lists Forte targets. That fallback is not verified as usable. Do not interpret plugin defaults as current availability. The source creates JSON jobs with type, backend and input; the service returns a job ID/status and requires polling/result retrieval. The plugin handles those contracts. No authentication or paid execution was attempted in this refresh.

Shots, reproducibility and operations

Use qml.set_shots(circuit, shots=1000) to produce a configured QNode and shots=None only on devices supporting analytic execution. A seeded simulator advances its RNG between calls; recreating the device reproduces the sequence. Do not globally cache one stateful device across independent concurrent experiments without understanding its state, random stream and backend thread safety.

Before hardware execution, budget parameter shifts, measurement groups, repetitions and shots, record the job identifier, and inspect failed/pending jobs before retrying. Blind resubmission can duplicate charged work. Provider queue/calibration and device topology drift require live checks. A matched local output establishes circuit semantics only, not hardware fidelity or access.

Primary sources

• Device API
• Lightning
• Qiskit remote device
• IBM runtime service
• Released Qiskit device source
• Braket plugin
• Cirq entry points
• Rigetti release
• Strawberry Fields plugin
• IonQ release
• IonQ v0.4 create-job contract