Features

Features#

PyTorch-native quantum training#

  • Trainable circuits as PyTorch modules: fq.Module exposes the quantum model to PyTorch, so loss.backward() and ordinary optimizers train quantum and classical layers in one loop.

  • Flat or named parameter groups: positional parameters and named groups such as {"encoder": (4,), "readout": ()} are both supported, with reproducible initialization and a module-local seed.

  • Stable training results: fq.train returns fq.TrainingResult with versioned summaries, and checkpoint and resume stay on fq.Module.

Several simulation representations, one program#

  • Statevector: the exact local path, on CPU or one GPU, with automatic dense reference reporting for small systems.

  • Matrix product state (MPS): low-entanglement systems at far larger wire counts than a dense statevector allows, including constrained TEBD.

  • Tensor networks: slicing and contraction for circuit structures where neither dense statevector nor MPS fits.

  • Optional JAX kernels behind the same PyTorch interface, with first-order gradient support across the bridge.

  • One command to switch: python examples/quick_start.py --mode sv|mps|tn runs the same hybrid model on three representations without editing code.

Compilation and export#

  • Target-independent optimization: flagquantum.compiler.optimize applies canonical rewrites to a fixed point and returns new IR.

  • Target-aware compilation: fq.compile(..., coupling_map=..., routing_strategy=...) emits only topology-valid two-qubit operations and records the routing decision.

  • Compiler plugins: independently installed packages may register a compiler through the extension registry, for example the QSteed plugin used for Quafu submission.

  • OpenQASM and framework export: circuit export to OpenQASM 3.0 and adapters for Qiskit, PennyLane, Cirq, Braket and CUDA-Q.

Measurement, noise and error correction#

  • Observables and outputs: fq.X, fq.Y, fq.Z with @ products and ordinary arithmetic build Hamiltonians; fq.expectation, fq.probabilities, fq.samples and fq.counts request results.

  • Backend-neutral noise: one NoiseModel drives exact density-matrix evolution, batched statevector trajectories and MPS quantum trajectories, with thermal relaxation, depolarizing, bit/phase flip and readout errors.

  • Hardware Pauli measurement planning: qubit-wise-commuting grouping with one sealed deployment package per group.

  • Dynamic circuits: mid-circuit measurement and classical feedback, with a read-only backend preflight.

  • Quantum error correction: a repetition-code memory experiment connecting syndrome extraction, decoding and correction.

Distributed and accelerator execution#

  • Sharded statevector training: one logical statevector across ranks, with distributed forward, backward and optimizer state.

  • Rank-owned MPS: distributed MPS forward, backward and optimizer state for workloads that do not fit one device.

  • PyTorch-native under a process group: the same module and result surface uses the sharded runtime once a multi-rank process group is initialized.

  • FlagOS accelerators through Torch-FL: the logical flagos:0 device is reached through Torch-FL, which owns vendor detection and dispatch.

Deployment, interop and ecosystem#

  • Sealed deployment packages: bind trained parameters, compile for a target and produce an auditable package before submission.

  • QPU digital twins: calibration-conditioned models that predict a device’s measurement distribution and carry explicit evidence boundaries.

  • Algorithm units: Grover search, amplitude estimation, quantum PCA, quantum k-medians, quantum kernel estimation and kernel ridge classification, feature selection as a QUBO, and QUBO-to-Ising mapping.

  • Circuit visualization: a text drawer for terminals and a Matplotlib drawer for publication figures.