OpenQML

PyPI · CI · Documentation

Trainable quantum circuits and tensor-network layers for PyTorch, JAX and Flax.

Installation

Requires Python 3.10 or later. Install the PyTorch integration to run the example below:

python -m pip install --upgrade "openqml[torch]"

Other options: openqml[jax], openqml[flax], or openqml for the NumPy core. The core depends on NumPy, packaging and psutil; training frameworks and quantum SDKs are optional.

Train a PyTorch model

import torch
from torch import nn
import openqml as oq

device = "cuda" if torch.cuda.is_available() else "cpu"
model = nn.Sequential(
    oq.torch.TensorNetworkAdapter(16, num_qubits=8, num_outputs=4, device=device),
    nn.Linear(16, 2, device=device),
)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)

x = torch.randn(32, 16, device=device)
target = torch.randint(0, 2, (32,), device=device)
optimizer.zero_grad()
logits = model(x)
loss = nn.functional.cross_entropy(logits, target)
loss.backward()
optimizer.step()
print(logits.shape)  # torch.Size([32, 2])

TensorNetworkAdapter(16, ...) preserves the last dimension: (..., 16) → (..., 16). Its MPO projections and quantum tree train together through ordinary autograd.

Choose an interface

These layers are available under oq.torch, oq.jax and oq.flax. JAX uses layer.init(key) and layer(parameters, inputs); Flax uses Linen init/apply. Dense import returns a module in PyTorch, (layer, parameters) in JAX, and (module, variables) in Flax.

Execution and documentation

Automatic circuit execution prefers a configured compatible QPU, then GPU, then CPU. Native tensor layers run on their selected framework device. There is no fixed circuit-width ceiling; memory, contraction width and work budgets determine capacity.

Tensor networks · Hybrid training · Quantum RL · Circuit conversion · QPU configuration

License

Apache-2.0.