Larger models

StochLift solves the extensive form directly. Three things help as models grow.

Parallel scenario solves

The wait-and-see solves, the evaluation of a fixed first stage in every scenario, and the out-of-sample test are independent per scenario. They run on a thread pool (the solvers release Python's lock while they work):

study = sl.lift(build_model, DATA, spec="uncertainty.yaml", n_jobs=8)

or --jobs 8 on the command line. The default is the number of CPUs, at most 8. The results do not depend on the number of threads. On a facility-location model with 20 sites, 60 customers and 50 scenarios, solve() went from 74 to 30 seconds.

Gurobi

study = sl.lift(build_model, DATA, spec="uncertainty.yaml", solver="gurobi")

or --solver gurobi. Every model, including the extensive form, is then solved with Gurobi (it needs gurobipy and a license that covers the model size).

Decomposition with mpi-sppy

For extensive forms too large to solve directly, hand the same scenarios to mpi-sppy, which has progressive hedging, Benders and other decomposition methods:

ex = study.to_mpisppy()      # all_scenario_names, scenario_creator, scenario_creator_kwargs

or write a module for mpisppy.generic_cylinders:

stochlift export model.py --spec uncertainty.yaml -o farmer_scen.py
python -m mpisppy.generic_cylinders --module-name farmer_scen --num-scens 100 --EF --EF-solver-name appsi_highs

The export works for models from any supported library: each scenario is rebuilt in Pyomo from the matrix StochLift read. The objective is in minimization form (negated for a maximization model). Decomposition needs mpi4py and an MPI installation; the extensive form does not. The test suite checks that mpi-sppy's extensive form reproduces StochLift's RP and first-stage decision. Mean-CVaR objectives are not exported (mpi-sppy has its own --cvar option).