PyTorch-first quantum computing
Build and train classical–quantum models with familiar PyTorch tools.
Explore simulation backends and run circuits on supported quantum hardware in one workflow.
$ pip install flagquantum$ git clone https://github.com/flagos-ai/FlagQuantum.git
Combine classical neural networks with trainable quantum circuits.
Build, optimize, and run hybrid models in a familiar PyTorch workflow.
Combine quantum circuits with PyTorch models, using familiar autograd and optimizers to train hybrid networks.
Choose statevector, MPS, or tensor-network simulation to match your circuit structure and available resources.
Keep circuits and measurements consistent as you switch between supported simulators and quantum hardware.
import torch
import flagquantum as fq
def circuit(parameters):
return fq.Circuit(2) \
.ry(0, parameters[0]) \
.cx(0, 1)
model = fq.Module(
circuit, n_parameters=1,
init=torch.tensor([0.25]),
)2 qubits · 1 trainable parameter
training = fq.train(
model,
optimizer=torch.optim.Adam(
model.parameters(), lr=0.08),
objective=lambda z: z.mean(),
steps=60,
)60 steps · Adam · learning rate 0.08
trained_circuit = circuit(
next(model.parameters()).detach())
measurement = fq.expectation(fq.Z(0))
# Local simulation
result = fq.run(
trained_circuit, outputs=measurement)
# Configured quantum hardware
hardware = fq.run(
trained_circuit, target="quafu:Baihua",
compiler="qsteed", outputs=measurement,
shots=1024,
)⟨Z₀⟩ = -0.996156The hardware call requires configured Quafu access; no remote job was run for this demo.
Turn a quantum circuit into a trainable module, ready to combine with classical neural networks.
import torch
import flagquantum as fq
def circuit(parameters):
return fq.Circuit(2) \
.ry(0, parameters[0]) \
.cx(0, 1)
model = fq.Module(
circuit, n_parameters=1,
init=torch.tensor([0.25]),
)2 qubits · 1 trainable parameter
Use automatic differentiation and familiar optimizers. Follow the loss as the model learns.
training = fq.train(
model,
optimizer=torch.optim.Adam(
model.parameters(), lr=0.08),
objective=lambda z: z.mean(),
steps=60,
)60 steps · Adam · learning rate 0.08
Keep your circuit and measurement explicit. Run locally or choose a configured quantum backend.
trained_circuit = circuit(
next(model.parameters()).detach())
measurement = fq.expectation(fq.Z(0))
# Local simulation
result = fq.run(
trained_circuit, outputs=measurement)
# Configured quantum hardware
hardware = fq.run(
trained_circuit, target="quafu:Baihua",
compiler="qsteed", outputs=measurement,
shots=1024,
)⟨Z₀⟩ = -0.996156The hardware call requires configured Quafu access; no remote job was run for this demo.
Compose independent modules—or bring classical control inside the quantum program.
Independent modules, joint optimization
Classical control within the circuit
Explore GPU simulation, the FlagOS multi-chip ecosystem, and quantum hardware through a shared development workflow.
GPGPU, domain-specific AI processors and heterogeneous backends for simulation and model training.
General-purpose GPU computing
Domain-specific AI computing, including MLU
Vendor-specific compute architectures
Connect circuits to quantum execution platforms.
Amazon Braket
Compile circuit → Submit job → Retrieve results
Chip names reflect the FlagOS / FlagTree backend ecosystem, not end-to-end FlagQuantum validation on every device. Architecture groups describe broad computing families, not identical microarchitectures; a vendor may span multiple product designs. Availability depends on the compiler branch, runtime, operators and workload. Quafu is a multi-route platform: superconducting access is connected, supercomputer simulation is being built jointly, and trapped-ion, neutral-atom and photonic access are coming soon. These are platform statuses; FlagQuantum execution depends on the selected backend and adapter. Quafu has an experimental FlagQuantum hardware adapter. Amazon Braket and IBM Quantum are listed with proposed OpenQASM / SDK and Qiskit integration paths; end-to-end integration remains to be verified. Circuit export alone does not establish hardware execution support.
FlagTree source ↗Learn from data, model molecules, optimize decisions, and understand quantum security.
Explore how a quantum layer changes learning, and compare it with a classical model.
10-class test accuracy: 92.2% hybrid · 96.2% classical
One fixed split · 4-qubit simulation
Study how molecular geometry and the choice of initial state affect variational optimization.
H₄ chain · 6 bond lengths · 3 initial states
Total energy · full-CI reference
Inspect quantum and classical candidates for the same weighted network problem.
20 nodes · exact best cut: 75
QAOA expectation · annealing mean
Follow RSA factorization and Grover search through closed, teaching-scale examples.
Shor + Grover · step-by-step walkthrough
Quantum algorithms · separate from AI
From device topology and control pulses to noise and validation: explore a computational counterpart of a quantum system.
A digital twin spans device connectivity, frequencies and couplings; Hamiltonian and pulse-level dynamics; noise, leakage and readout; and calibration-driven validation. The view illustrates these layers without claiming an implemented end-to-end FlagQuantum digital twin. Device-specific data and independent checks are essential.
Read the research ↗Swipe to explore the full model ↔
Map qubits, coupling, frequencies and device geometry into a computational model. The nine-node layout is illustrative, not a replica of a connected processor.
Inspect a rotated surface-code patch, follow detection events through time, and connect decoding to logical resources.
This patch encodes one logical qubit. Logical X and Z can be represented by boundary-to-boundary strings. Fault-tolerant computation also requires repeated measurement, a noise-aware decoder, logical operations such as lattice surgery, and a resource budget. Larger distance helps only in a suitable noise regime; no threshold or logical error rate is estimated here.
Surface-code reference ↗Swipe to explore the full model ↔
Distance-3 rotated surface code: 9 data qubits, 4 X checks and 4 Z checks. Colored faces and boundary checks show stabilizer support. One ancilla per check gives 17 physical qubits.
Install FlagQuantum in your Python environment with pip.
$ pip install flagquantumClone the full source and follow the setup instructions in the repository.
$ git clone https://github.com/flagos-ai/FlagQuantum.gitContribute models, build tools, and share experiments with FlagQuantum. Bring your ideas to classical–quantum computing, from research to applications.
Institutions & platforms
FlagOS · Beijing Academy of Artificial Intelligence (BAAI)
Quafu Quantum Cloud · Beijing Academy of Quantum Information Sciences (BAQIS)