torch-quantum

🌀 torch-quantum

Quantum-native deep-learning layers that slot directly into PyTorch and run on NVIDIA CUDA-Q simulators or real QPUs.

Why use torch-quantum?

Current quantum SDKs expose powerful primitives but leave many “deep-learning conveniences” to the user. Torch-Q fills that gap by giving PyTorch practitioners a familiar module interface, autograd-compatible parameter-shift gradients and ready-made feature-map / ansatz building blocks. You build models exactly as you do with classic layers, choose a CUDA-Q target (CPU, GPU or cloud QPU) and start iterating.

Classical DL	torch-quantum	Notes
`nn.Linear`	`quantum.QuantumLayer`	drop-in layer that outputs class probabilities
GPU accel	cuStateVec / cuTensorNet	automatic, single or multi-GPU
Autograd	parameter-shift implemented in pure PyTorch	works with all optimizers

Features

Drop-in PyTorch modules (QuantumLayer, QNN, HybridQNN) that behave like any other nn.Module and are compatible with torch.optim.
Flexible circuit construction: pick standard feature maps (Z, ZZ) or supply custom kernels; choose from RealAmplitudes, EfficientSU2, PauliTwoDesign ansätze or roll your own.
Hybrid workflow: seamlessly chain classical layers before or after a quantum block, making it easy to build quantum-enhanced CNNs, MLPs or Transformers.
Multiple back-ends: one line (cudaq.set_target(...)) switches from local CPU simulation to GPU acceleration or a cloud device (IonQ, Quantinuum, OQC, Infleqtion, Pasqal, QuEra…).
Exact gradients via the parameter-shift rule, automatically invoked by loss.backward(); no manual circuit plumbing required.
Research-friendly utilities: empirical Fisher information, layer-wise kernel drawers, circuit depth counters and parameter initialisation helpers.

Installation

python -m pip install --upgrade pip
pip install cudaq
git clone https://github.com/SeroviICAI/cuda-quantum.git
python - <<'PY'
import importlib.util, pathlib, os, shutil
src = pathlib.Path("cuda-quantum/python/cudaq/kernel/quake_value.py").resolve()
dst = pathlib.Path(importlib.util.find_spec("cudaq.kernel.quake_value").origin)
dst.unlink()
os.symlink(src, dst)
PY
pip install torch-qu

Quick example (4-qubit classifier)

import torch, torch.nn as nn, torch.optim as optim
from torchq.models import QNN
import cudaq

# Use fast GPU simulator if available
cudaq.set_target("nvidia", option="fp32")

model = QNN(
    in_features = 4,        # qubits / input dimension
    out_features = 3,       # number of classes
    num_layers = 2,         # ansatz depth
    shots = 1024,           # measurement shots per forward pass
    feature_map = "zz",     # entangling data embedding
    var_form = "efficientSU2",
    reupload = False
)

x = torch.randn(16, 4)              # mini-batch
y = torch.randint(0, 3, (16,))      # labels
opt = optim.Adam(model.parameters(), lr=0.02)
criterion = nn.CrossEntropyLoss()

for step in range(50):
    opt.zero_grad()
    logits = model(x)               # parameter-shift handled automatically
    loss = criterion(logits, y)
    loss.backward()
    opt.step()
print("final loss:", loss.item())

Change back-end to a real device with one line:

# Add CREDENTIALS here
cudaq.set_target("ionq", qpu="qpu.aria-1")  # 25-qubit trapped-ion hardware

All circuits are now queued to the cloud without further code changes.

Documentation & book

The open-access book “Toward a Quantum Advantage in Deep Learning Architectures” lives in docs/ and is rendered online at https://SeroviICAI.github.io/torch-quantum/book. It introduces quantum mechanics, CUDA-Q programming, variational circuits and information-geometric capacity in detail and Torch-Q code examples.

Citation

@software{torchquantum2025,
  author  = {Sergio Rodríguez Vidal},
  title   = {torch-quantum: Quantum-ready layers for PyTorch},
  year    = {2025},
  url     = {https://github.com/SeroviICAI/torch-quantum},
  license = {Apache-2.0}
}

License

Apache 2.0 — free to use, modify and distribute, with permissive terms.

Happy hacking — and welcome to quantum-enhanced deep learning!

This site is open source. Improve this page.