Maestro
Maestro is Qoro’s quantum circuit simulator. A C++ engine sits behind one API and offers many simulation methods, CPU and GPU execution, noise models, and native expectation values. It runs in Divi on a laptop, as Qoro’s cloud simulator fleet, and on HPC clusters.
๐ฆ GitHub: QoroQuantum/maestro ย |ย ๐ Reference docs: qoroquantum.github.io/maestro ย |ย
pip install qoro-maestro
The CPU engine, with every simulation method and the Qiskit and PennyLane plugins, is open source under GPL-3.0. GPU acceleration is a separately licensed plugin for your own hardware, available in Maestro Enterprise and free for academic use. On Qoro’s cloud, GPU simulation is open to every account with no licence. Pick the GPU simulator cluster for the job (see Deployment Options).
Architecture
flowchart TD
A["Python ยท C/C++ API ยท JSON request API ยท CLI<br/>Qiskit and PennyLane plugins"] --> B["Maestro core<br/>circuit IR ยท optimisation passes ยท noise models"]
B --> C["QCSim<br/>(CPU, all methods)"]
B --> D["GPU plugin<br/>(CUDA / cuQuantum)"]
B --> E["Distributed GPU<br/>(multi-GPU, MPI)"]
B --> F["Qiskit Aer ยท QuEST<br/>(optional engines)"]
The core parses circuits (its own circuit model, or OpenQASM), runs optimisation passes such as gate fusion, causal-cone reduction, and observable clustering, and hands the circuit to a simulator engine. GPU, distributed, and QuEST engines are runtime plugins loaded only when present, so the pip wheel stays CPU-only and lightweight.
Simulation methods
| Method | Use it for | CPU | GPU |
|---|---|---|---|
| Statevector | Exact simulation of any circuit up to about 30 qubits per node, more with distributed GPUs | โ | โ |
| Matrix product state (MPS) | Many qubits with limited entanglement; accuracy set by bond dimension | โ | โ |
| Stabilizer | Clifford circuits of any size | โ | |
| Extended stabilizer | Mostly-Clifford circuits with a few non-Clifford gates | โ | |
| Pauli propagation | Expectation values on circuits too wide for a statevector | โ | โ |
| Tensor network | Contraction-based simulation of structured circuits | โ | โ |
| Density matrix | Exact mixed-state noise on small registers | โ | โ |
| Matrix product operator (MPO) | Exact noise channels on larger, low-entanglement registers | โ | โ |
| Path integral | Single-amplitude and probability queries | โ |
A composite (p-block) mode splits a circuit into independent qubit blocks and simulates each block on its own statevector. It also underpins Maestro’s distributed-QC simulations.
Maestro Auto
Choosing the right method per circuit can mean orders-of-magnitude differences in runtime, especially across a heterogeneous batch. Maestro Auto makes that choice automatically. It uses runtime estimators benchmarked on the real engines (see the Composer engine) to predict the cost of each candidate simulator and method, then picks the fastest one that is accurate enough. Clifford circuits go to the stabilizer engine, low-entanglement circuits to MPS, and dense circuits to statevector.
Maestro Auto is on by default for jobs on Qoro’s cloud (Solo and Dedicato). It can also be requested through Maestro’s native JSON request API. A local pip install defaults to the CPU statevector, and you pick a method explicitly with SimulatorConfig.
Beyond sampling
- Expectation values of Pauli observables are computed natively, without measurement circuits. This is what lets Qoro’s cloud return expectation values directly to Divi.
- Noise models cover Pauli, coherent, relaxation, readout, crosstalk, and time-correlated noise, either sampled as trajectories or applied exactly on density-matrix and MPO states. See Noise Models.
- Gradients and state access include adjoint differentiation, statevectors, probabilities, inner products, and mirror fidelity for checking MPS truncation.
- Error correction. A Stim/Sinter integration lets QEC experiments sample through Maestro, for example with coherent noise that Stim cannot represent.
Ways to use it
| Interface | For |
|---|---|
Python (qoro-maestro) |
Scripts and notebooks; Divi’s MaestroSimulator |
| qiskit-maestro | A Qiskit BackendV2, Sampler, and Estimator; a drop-in replacement for Aer |
| pennylane-maestro | A PennyLane device with adjoint gradients |
| C/C++ library and JSON request API | Embedding in other software and services |
maestro CLI |
Running OpenQASM files from the shell or batch scripts |
| Local server and QRMI | Multi-user HPC sites behind Slurm (see HPC & Deployment) |
| Qoro cloud | Divi’s QoroService or the REST API (see Composer) |