QQA4CO

Try QQA in your browser

The hosted Streamlit studio needs no local installation. Build a problem, run QQA, inspect convergence and feasibility, and compare supported methods from a browser. For a local studio, install qqa[gui] and run qqa gui.

QQA4CO is a GPU-first primal-search and hybrid optimisation runtime. Quasi-Quantum Annealing (QQA) generates diverse candidates, structure-aware repair and local search refine them, and optional mathematical solvers certify them when requested.

The default qqa.solve(...) route is pure QQA. Exact solvers, GNNs, dashboard dependencies, and public benchmark parsers are explicit extras.

Install

Core CPU/GPU solver:

python -m pip install --upgrade pip
python -m pip install --upgrade qqa

Common optional installations:

python -m pip install --upgrade "qqa[gui]"       # Streamlit studio
python -m pip install --upgrade "qqa[benchmark]" # MIPLIB/QPLIB + SCIP
python -m pip install --upgrade "qqa[highs]"     # HiGHS LP/MIP adapter
python -m pip install --upgrade "qqa[cpsat]"     # OR-Tools CP-SAT adapter
python -m pip install --upgrade "qqa[service]"   # schema-only FastAPI job service
python -m pip install --upgrade "qqa[pignn]"     # experimental CRA/CPRA GNN
python -m pip install --upgrade "qqa[triton]"    # optional fused CUDA kernels
python -m pip install --upgrade "qqa[dev]"       # tests, lint, typing, docs

QQA4CO supports CPython 3.10--3.14. PyPy is not currently supported because the solver depends on PyTorch, whose official distributions target CPython.

PyTorch chooses CPU or CUDA according to the installed wheel. Follow the GPU setup guide when a specific CUDA build is required.

Quickstart

import networkx as nx
import qqa

graph = nx.random_regular_graph(d=3, n=100, seed=0)
problem = qqa.MaximumIndependentSet(graph, penalty=2.0)

result = qqa.solve(
    problem,
    goal="best",
    budget="30s",
    device="auto",
    seed=0,
)

print(result.status.value)
print(result.best_obj)      # original mathematical objective
print(result.feasible)
print(result.runtime)

The high-level result does not hide transformations:

print(result.objective_value)
print(result.internal_energy)
print(result.merit_value)
print(result.raw_solution)
print(result.repaired_solution)
print(result.violations.maximum_violation)
print(result.plan.explain())

Legacy research calls such as qqa.anneal, qqa.simulated_annealing, qqa.population_annealing, and qqa.discrete_langevin remain available. New integrations should prefer solve, plan, and inspect.

The one-call goal can be best, feasible, prove, diverse, or pareto. Durations accept ms, s, m, and h. prove enables an exact route but still returns a proof status only when the backend certifies the original model.

Inspect and plan before solving

inspection = qqa.inspect(problem)
print(inspection.to_dict())

doctor = qqa.doctor(problem, replicas=128)
print(doctor.explain())

plan = qqa.plan(problem, profile="quality", device="cuda")
print(plan.explain())

The planner reports domains, sparse factors, connected components, selected engine, refinements, certification route, VRAM estimate, replica count, and the reason for each choice. It does not execute the solver.

The Model Doctor additionally checks bounds, factor capabilities, scaling, curvature, presolve contradictions, decomposition, and route/proof support. Missing real or integer bounds are never silently replaced by a guessed box.

Profiles

qqa.solve(problem, profile="fast")
qqa.solve(problem, profile="balanced")
qqa.solve(problem, profile="quality")
qqa.solve(problem, profile="reproducible", seed=7)

Additional profiles:

Checkpoint, events, and portable packages

Long QQA runs can be resumed without pickle payloads:

qqa.solve(
    problem,
    budget="10m",
    checkpoint_path="run.qqacp",
    checkpoint_interval=100,
)
result = qqa.solve(problem, budget="20m", resume_from="run.qqacp")

The atomic checkpoint verifies model/config fingerprints, tensor checksums, optimizer and RNG state, schedule state, incumbent, population, and historical archive. Paths stay outside result provenance. SolveResult.events uses the versioned schema-v2 stream, and qqa.runtime.export_result_package writes a checksum-protected, environment-neutral exchange bundle.

Python source and pickle inputs are trusted-local features and are denied by default. Use trusted=True in Python, --allow-unsafe-python in the CLI, or QQA_ALLOW_CUSTOM=1 for a single-user local GUI. Do not enable custom code in a shared deployment.

The Colab-ready 13_typed_primal_dual_runtime.ipynb walks through diagnosis, solve events, the cockpit, checkpoint/resume, and a verified package without credentials or environment-specific paths.

Advanced configuration is strict. Unknown or misspelled options raise an error:

config = qqa.SolverConfig.for_profile(
    "quality",
    replicas=512,
    epochs=4000,
    schedule="reheat",
    compile_core=True,
    exact_backend="none",
)
result = qqa.solve(problem, config=config)

Model inputs

The common factor-based ModelIR represents binary, spin, integer, real, categorical, and permutation blocks. Native factors include:

• linear, sparse quadratic, and higher-order terms;
• clauses, cardinality, all-different, assignment, Potts, and table terms;
• indicators, SOS1/SOS2, piecewise-linear and logical relations;
• precedence, no-overlap, cumulative resource, flow, matching, and subtour terms;
• scenario mean/worst-case/CVaR and chance-constraint aggregation.

Supported file routes:

qqa inspect model.mps
qqa plan model.mps --profile quality --device auto
qqa solve model.mps --profile balanced --device auto

MIPLIB and QPLIB

Install the benchmark extra once:

pip install "qqa[benchmark]"

Fetch and inspect one public instance:

qqa benchmark fetch miplib --instance air05 --output benchmarks/miplib
qqa benchmark inspect benchmarks/miplib/air05.mps.gz

qqa benchmark fetch qplib --instance QPLIB_0031 --output benchmarks/qplib
qqa benchmark inspect benchmarks/qplib/QPLIB_0031.qplib

Run QQA-centred SG-CQQA or a paired baseline comparison:

qqa benchmark run INSTANCE --solver sg-cqqa --time-limit 60 --device auto

qqa benchmark compare INSTANCE \
  --solvers scip-aggressive sg-cqqa \
  --seeds 0 1 2 \
  --time-limit 60 \
  --output results.json

The campaign runner uses equal wall-clock budgets, paired seeds, checkpointing, direction-aware objectives, feasibility, bounds, gaps, time to first feasible, and primal-integral metrics. See the MIPLIB/QPLIB guide and the transparent published result summary.

Optional exact completion

Pure QQA remains the default:

result = qqa.solve(model, profile="balanced", exact_backend="none")

Certification is explicit:

result = qqa.solve(model, profile="certify", exact_backend="scip", budget=60)
print(result.best_bound, result.relative_gap, result.proven_optimal)

SCIP supports the broadest mixed nonlinear route. HiGHS is available for sparse linear LP/MIP models, and CP-SAT for bounded integral linear models with integral coefficients. Unsupported semantics fail clearly; adapters never silently round, drop, or reinterpret constraints.

Structured optimisation

problem = qqa.TSP(N=30, seed=0)  # Sinkhorn permutation relaxation
result = qqa.solve(problem, profile="quality")

print(result.raw_solution)       # untouched optimiser output
print(result.repaired_solution)  # Hungarian projection + 2-opt, when changed

Mixed-variable models use scaled constraints, per-constraint augmented Lagrangian multipliers, feasibility-first archives, and explicit repair. Sparse graph QUBOs use edge factors and incremental one-flip local search instead of an N × N matrix in the hot path.

Black-box optimisation

problem = qqa.BlackBoxProblem(
    [qqa.Real("x", -2.0, 2.0), qqa.Integer("n", 0, 8)],
    lambda point: (point["x"] - 0.4) ** 2 + (point["n"] - 3) ** 2,
)

result = problem.solve(
    budget=100,
    workers=4,
    trust_regions=3,
    evaluation_database="evaluations.sqlite",
)

The optional evaluation database caches encoded points and records pending/running/completed/failed/timed-out/cancelled states. Large campaigns can select random-Fourier-feature surrogates; discrete batch acquisition can opt into QQA with acquisition_optimizer="qqa".

Feature status

Documentation

• Quickstart
• Architecture
• Advanced opt-in runtime
• Algorithm
• Problem catalog
• Mixed optimisation
• Benchmark protocol
• CLI reference
• Python API
• Migration guide

Development

git clone https://github.com/Yuma-Ichikawa/QQA4CO.git
cd QQA4CO
python -m pip install --upgrade pip
python -m pip install --upgrade -e ".[dev]"
ruff check src tests scripts app
ruff format --check src tests scripts app
pytest -q
mkdocs build --clean --strict

Large datasets and generated campaign trajectories are not source files. Keep only tiny smoke instances, checksums, licenses, and compact summaries in the repository; use the documented fetch commands and release/CI artifacts for full data.

Research papers and BibTeX

QQA4CO builds on three peer-reviewed continuous-relaxation methods. The default solver is pure QQA; the CRA/CPRA-inspired graph-learning route is an explicit pignn extra.

QQA / PQQA — ICLR 2025

Paper (OpenReview) · arXiv

@inproceedings{ichikawa2025optimization,
  title={Optimization by Parallel Quasi-Quantum Annealing with Gradient-Based Sampling},
  author={Ichikawa, Yuma and Arai, Yamato},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=9EfBeXaXf0}
}

CRA — NeurIPS 2024

Paper (NeurIPS Proceedings) · OpenReview · Reference implementation

@inproceedings{NEURIPS2024_54191f42,
  author={Ichikawa, Yuma},
  booktitle={Advances in Neural Information Processing Systems},
  editor={A. Globerson and L. Mackey and D. Belgrave and A. Fan and U. Paquet and J. Tomczak and C. Zhang},
  pages={47189--47216},
  publisher={Curran Associates, Inc.},
  title={Controlling Continuous Relaxation for Combinatorial Optimization},
  url={https://proceedings.neurips.cc/paper_files/paper/2024/file/54191f424e9013fc1d7b923f6e45dff4-Paper-Conference.pdf},
  volume={37},
  year={2024},
  doi={10.52202/079017-1495}
}

CPRA — TMLR 2025

Paper (OpenReview / TMLR) · Reference implementation

@article{ichikawa2025continuous,
  title={Continuous Parallel Relaxation for Finding Diverse Solutions in Combinatorial Optimization Problems},
  author={Ichikawa, Yuma and Iwashita, Hiroaki},
  journal={Transactions on Machine Learning Research},
  issn={2835-8856},
  year={2025},
  url={https://openreview.net/forum?id=ix33zd5zCw}
}

License

QQA4CO is distributed under the BSD 3-Clause License.