Offload Qiskit workloads

Install the qiskit extra to use ibm.qdmi.offloader, ibm-sampler, and ibm-estimator. sample() returns joint measurement counts. estimate() runs VQE with L-BFGS-B and returns the complete Qiskit VQEResult.

Local simulation

Use local=True, simulator=True to run without Slurm or IBM access:

The documentation executes these local simulator cells and displays their results. Slurm and hardware examples below remain unexecuted.

from qiskit import QuantumCircuit

from ibm.qdmi.offloader import sample

circuit = QuantumCircuit(2)
circuit.h(0)
circuit.cx(0, 1)
circuit.measure_all()
counts = sample(circuit, shots=128, local=True, simulator=True)
assert sum(counts.values()) == 128
assert set(counts) <= {"00", "11"}
counts

{'00': 59, '11': 69}

The estimator offloader returns a full VQE result. This short run demonstrates the API; three optimizer iterations do not guarantee convergence to the ground state of the Pauli-Z observable.

import math

from qiskit.circuit import Parameter
from qiskit.quantum_info import SparsePauliOp
from qiskit_algorithms.utils import algorithm_globals

from ibm.qdmi.offloader import estimate

algorithm_globals.random_seed = 7
ansatz = QuantumCircuit(1)
ansatz.ry(Parameter("theta"), 0)
result = estimate(ansatz, SparsePauliOp("Z"), maxiter=3, local=True, simulator=True)
energy = float(result.eigenvalue.real)
assert math.isfinite(energy) and -1.0 <= energy <= 1.0
assert result.optimal_parameters is not None
{"energy": energy, "optimizer_evaluations": result.optimizer_evals}

{'energy': -0.0009765625, 'optimizer_evaluations': None}

The simulator uses MQT Core’s DDSIM device. Sampling preserves joint classical register order. VQE transpiles the parameterized ansatz and applies its layout to the observable before optimization. Local and worker execution share this path.

Slurm execution

Omit local=True to submit one blocking srun task:

counts = sample(circuit, shots=128, simulator=True, partition="quantum", timeout=300)

Install the same package and compatible Python, Qiskit, and Qiskit Algorithms versions on the submitting host and worker. Both need access to IBM_JOBS_DIR, which defaults to ~/.qdmi_jobs. Use a private directory on a shared filesystem. The worker exchanges QPY inputs and base64-encoded pickled results. Estimation also uses a pickled observable. Only load files and results from trusted workers; pickle can execute code. Successful calls remove their inputs. Failed calls retain their per-call directory for diagnosis.

partition overrides IBM_SLURM_PARTITION, then defaults to quantum. licenses forwards Slurm’s name[:count][,name[:count]...] request. nodes defaults to one; the worker always uses one task. The timeout limits the srun process wait. It does not guarantee cancellation of an IBM job already submitted by the worker.

IBM execution

Set simulator=False only for an authorized hardware workload. This is the API default. Supply IBM_QUANTUM_API_KEY and IBM_QUANTUM_INSTANCE_CRN through the worker environment. Configure their propagation according to site policy. Credentials never appear in worker arguments or serialized inputs.

backend_name applies to both local and offloaded calls and overrides IBM_QUANTUM_BACKEND. Slurm passes it directly as the worker’s --backend-name argument. No scheduler plugin is required. See backend selection for native configuration.

Each submitted circuit retains the native 60-second QPU execution cap. VQE can submit many circuits per optimizer iteration; maxiter does not cap shots, circuit submissions, or total spending. Agree on a workload budget before using hardware. See job contracts for cancellation.

Workers also run directly with trusted serialized inputs:

ibm-sampler circuit.qpy --shots 128 --simulator
ibm-estimator ansatz.qpy operator.pkl --maxiter 3 --simulator

Their standard output carries the serialized result for the calling offloader.