lp-ip-craft

Linear programming (LP) and mixed-integer linear programming (MILP) in Python, solved by a native Rust engine: a bounded-variable dual revised simplex on a sparse LU factorization, presolve, cuts and branch-and-bound. Every returned point is evaluated independently of the solver (objective, row activities, bound and integrality violations), and lp_residuals checks LP duals.

Install

python -m pip install lp-ip-craft

Wheels (abi3, CPython 3.10 or newer) cover Linux x86_64, Windows x64 and macOS x86_64/arm64 and need no compiler. On other platforms pip builds the source distribution, which needs Rust 1.85 or newer. The package depends on nothing else at run time.

Example

from lp_ip_craft import Model, solve

# maximize 5x + 4y  subject to  6x + 4y <= 24,  x + 2y <= 6,  0 <= x, y <= 10, integer
model = Model(
    objective=[5.0, 4.0], sense="maximize",
    lower=[0.0, 0.0], upper=[10.0, 10.0],
    row_lower=[float("-inf"), float("-inf")], row_upper=[24.0, 6.0],
    indptr=[0, 2, 4], indices=[0, 1, 0, 1], values=[6.0, 4.0, 1.0, 2.0],  # A in CSR form
    kinds=["integer", "integer"],
)
result = solve(model, backend="native")
print(result.status, result.candidate, result.evaluation.objective, result.best_bound)
# optimal [4.0, 0.0] 20.0 20.0

Pass backend="native": the default backend="dummy" only evaluates a start point. Options sets the tolerances and the numerical width (threads, 1 to 8); NativeOptions holds the engine's controls, including time, node and iteration limits, and rejects unknown names. solve_batch solves independent models with up to 16 workers. evaluate and lp_residuals check points and duals from any solver.

Limits

• A MIP needs finite bounds on all variables, continuous ones included, as given or derived by root inference (which runs with presolve on).
• Linear objectives only: no QP or MIQP.
• optimal and the gap are numerical findings under the stated tolerances. For hard industrial MIPs, compare against a mature solver such as HiGHS.

More

• Repository and the source build; the Rust crates are lp-ip-craft and lp-ip-craft-native
• AI agent guide: modelling, options, results, MPS import and the job API
• Changelog
• Shared solver-selection guide and ecosystem overview: when black-box optimization (fcmaes) or SQP (SQP-craft) fits better, and durable shell/MCP jobs through the lp-ip-job-worker (built from source)

MIT license.