TorQ-bench 0.1.3

Tensor Operations for Research in Quantum systems - Benchmarking.

TorQ-bench

Benchmarks and PennyLane comparisons for TorQ.

TorQ-bench is a small companion package that lets you run the same layer logic in TorQ and PennyLane to compare outputs and timing. It intentionally keeps PennyLane out of the core TorQ package.

Developed while writing the paper: Quantum Physics-Informed Neural Networks for Maxwell’s Equations: Circuit Design, “Black Hole” Barren Plateaus Mitigation, and GPU Acceleration

by Ziv Chen, Gal G. Shaviner, Hemanth Chandravamsi, Shimon Pisnoy, Steven H. Frankel, Uzi Pereg

https://arxiv.org/abs/2506.23246

For the simpler version associated with the following paper, please use the simple_for_PINNACLE branch: PINNACLE: An Open-Source Computational Framework for Classical and Quantum PINNs

by Shimon Pisnoy, Hemanth Chandravamsi, Ziv Chen, Aaron Goldgewert, Gal Shaviner, Boris Shragner, Steven H. Frankel

https://arxiv.org/abs/2604.15645

It is compatible with TorQ simple_for_PINNACLE branch and PINNACLE repos. Also, the simple_for_PINNACLE branch is the same as torq-bench v0.1.2 on PyPI.

Install

From source (recommended while developing):

pip install -e .[pennylane]

Or from PyPI:

pip install torq-bench[pennylane]

TorQ-bench depends on torq-quantum>=0.1.3. PennyLane is optional and only required for the comparison wrappers.

Note: the PyPI distribution is torq-quantum, while the Python import package remains torq.

Quickstart: compare TorQ vs PennyLane

import torch
from torq.QLayer import QLayer
from torq_bench import PennyLaneQLayer

n_qubits = 4
n_layers = 2
x = torch.rand(8, n_qubits)

torq_layer = QLayer(n_qubits=n_qubits, n_layers=n_layers)
pl_layer = PennyLaneQLayer(
    n_qubits=n_qubits,
    n_layers=n_layers,
    pennylane_dev_name="default.qubit",
)

y_torq = torq_layer(x)
y_pl = pl_layer(x)

Notes:

• PennyLaneQLayer supports TorQ ansatz names: basic_entangling, single_rot_basic_ent, strongly_entangling, cross_mesh, cross_mesh_2_rots, cross_mesh_cx_rot, tile, and no_entanglement_ansatz.
• data_reupload_every is supported.
• Config-driven ansatz options are forwarded to PennyLane: single_rot_basic_ent uses single_rotation_gate, and tile uses tile_rotation_params, single_rotation_gate, tile_sublayers, and tile_cyclic. The tile_cyclic wrap-around CNOT is applied only when n_qubits > 2 and even, matching TorQ.
• Output observables follow TorQ's current config.observables API: None for per-qubit Z, Pauli strings like "X" or "XI_ZZ", or Hermitian matrices with shape [2,2], [n_qubits,2,2], [2**n,2**n], or [m,2**n,2**n].

Using PennyLaneComparison directly

import torch
from torq_bench import PennyLaneComparison

n_qubits = 4
n_layers = 2
weights = torch.rand(n_layers, 1, n_qubits, 3)  # current TorQ no-data layout
x = torch.rand(n_qubits)

qc = PennyLaneComparison(n_qubits=n_qubits, n_layers=n_layers, weights=weights)
circuit = qc.circuit_strongly_entangling()
state = circuit(x)

PennyLaneComparison also accepts the older no-data shape [n_layers, *per_layer_shape] for backward compatibility.

Run the built-in demo

The comparison module includes a demo that builds and draws the supported TorQ ansatzes, including single_rot_basic_ent and tile variants. It uses qml.draw_mpl, so you will need matplotlib installed.

python -m torq_bench.PennyLaneComparison

License

MIT