NIRYUKTI — sparse optimization for Python

An independently implemented C++20 optimization engine with a Python model API, native sessions, a command-line interface, authenticated HTTP service and offline HTML reporting. No existing optimization solver is used to solve your model.

Install

pip install niryukti
niryukti devices

Python 3.10+ is required. Linux x86_64 wheels bundle the FP64 CPU executable and native library. Source installations require CMake 3.24+ and a C++20 compiler. CUDA is optional and is not bundled in CPU wheels. Version 0.2.1 includes the service and report additions described below.

Build and solve a model

from niryukti import Model, NativeSession

model = Model()
model.add_var("x", ub=10)
model.add_constraint({"x": 1}, ">=", 2)
model.set_objective({"x": 3})
result = model.solve(device="auto", method="auto", time_limit=60)
print(result["status"], result["objective"])

# Keep model data in a native session for repeated requests.
with NativeSession(model) as session:
    print(session.solve())

LP, supported convex sparse QP, MILP and convex MIQP are available. Sparse convexity certification and advanced integer algorithms have documented limits; this is a research optimizer, not a validated replacement for every industrial solver. Nonconvex global optimization is unsupported. Always inspect status and verification diagnostics rather than assuming a returned point is optimal.

CLI, verification and reports

niryukti inspect model.mps
niryukti solve model.mps --method auto --device auto --time-limit 60 --json-out result.json
niryukti verify model.mps result.json
niryukti report result.json --output report.html

Automatic selection considers structure and GPU memory; explicit methods remain available. Reports are portable, printable HTML derived from saved telemetry. Rendering a report does not independently certify its input.

HTTP API

export NIRYUKTI_API_TOKEN="your-long-random-secret"
niryukti serve --host 127.0.0.1 --port 8090 --workers 2 --max-time 300

Send Authorization: Bearer <token> to /v1/health, /v1/solve or /v1/report. Solve requests contain a native JSON model and optional solver settings. The service bounds request size, concurrency and solve time, and does not accept client executable paths. It is synchronous; disconnecting does not cancel computation. Put remote deployments behind HTTPS and appropriate access controls. Full examples are in the repository's docs/api.md.

CUDA and source builds

Build the repository with -DNIRYUKTI_CUDA=ON, then set NIRYUKTI_BINARY to its executable and NIRYUKTI_LIBRARY to libniryukti_c.so when using native sessions. Legacy VANTAGE environment variables remain compatibility aliases. Build distributions with python -m build; CPU wheels do not download GPU code.

License

AGPL-3.0-only for original project code, with third-party notices retained. Copying and commercial use are permitted under the license terms. Covered modified distributions retain AGPL; modified network deployments must offer corresponding source to users. See the bundled LICENSE and NOTICE.

Version 0.2.2

Adds checkpoint/resume for simplex, barrier and concurrent portfolios, guarded sparse singular-PSD ordering, general-integer bound conflicts and integer-lattice cuts. CUDA source builds also include continuous bound derivation replay, dual reconstruction and GPU CSR compaction. CPU distributions include the same mathematical core, but require a separate CUDA build for GPU execution.

Integer incumbent verification is distinct from replaying a full search-tree proof. Large/difficult PSD recognition, general dual conflict analysis and direct GPU barrier factorization remain restricted. No universal speedup is claimed.