CuOpt Solver Integration (#839)

Iroy30 · web-flow · commit 0d6e0a5400b0 · 2025-05-29T11:20:13.000+02:00
* add cuopt solver interface

* fix mip

* update constraint extraction and tests

* incorporate review changes

* Update cuopt_api.py

* Delete pulp/tests/test_pulp_temp.py

* run mypy

* Update cuopt_api.py
diff --git a/pulp/apis/__init__.py b/pulp/apis/__init__.py
@@ -13,6 +13,7 @@
 from .sas_api import *
 from .scip_api import *
 from .xpress_api import *
+from .cuopt_api import *
 
 _all_solvers: List[Type[LpSolver]] = [
     CYLP,
@@ -41,6 +42,7 @@
     COPT_CMD,
     SAS94,
     SASCAS,
+    CUOPT,
 ]
 
 import json
diff --git a/pulp/apis/cuopt_api.py b/pulp/apis/cuopt_api.py
@@ -0,0 +1,277 @@
+import ctypes
+import os
+import subprocess
+import sys
+import warnings
+from uuid import uuid4
+from ..constants import (
+    LpBinary,
+    LpConstraintEQ,
+    LpConstraintGE,
+    LpConstraintLE,
+    LpContinuous,
+    LpInteger,
+    LpMaximize,
+    LpMinimize,
+    LpStatusInfeasible,
+    LpStatusNotSolved,
+    LpStatusOptimal,
+    LpStatusUnbounded,
+    LpStatusUndefined,
+)
+from .core import (
+    LpSolver,
+    LpSolver_CMD,
+    PulpSolverError,
+    clock,
+    ctypesArrayFill,
+    sparse,
+)
+
+# Constraint Sense Converter
+sense_conv = {
+    LpConstraintLE: "L",
+    LpConstraintGE: "G",
+    LpConstraintEQ: "E",
+}
+
+
+class CUOPT(LpSolver):
+    """
+    The CUOPT Optimizer via its python interface
+    """
+
+    name = "CUOPT"
+
+    try:
+        global cuopt
+        import cuopt  # type: ignore[import-not-found, import-untyped, unused-ignore]
+
+        global np
+        import numpy as np  # type: ignore[import-not-found, import-untyped, unused-ignore]
+    except:
+
+        def available(self):
+            """True if the solver is available"""
+            return False
+
+        def actualSolve(self, lp, callback=None):
+            """Solve a well formulated lp problem"""
+            raise PulpSolverError("CUOPT: Not available")
+
+    else:
+
+        def __init__(
+            self,
+            mip=True,
+            msg=True,
+            timeLimit=None,
+            gapRel=None,
+            warmStart=False,
+            logPath=None,
+            **solverParams,
+        ):
+            """
+            :param bool mip: if False, assume LP even if integer variables
+            :param bool msg: if False, no log is shown
+            :param float timeLimit: maximum time for solver (in seconds)
+            :param float gapRel: relative gap tolerance for the solver to stop (in fraction)
+            :param bool warmStart: if True, the solver will use the current value of variables as a start
+            :param str logPath: path to the log file
+            :param solverParams: solver setting paramters for cuopt
+            """
+
+            LpSolver.__init__(
+                self,
+                mip=mip,
+                msg=msg,
+                timeLimit=timeLimit,
+                gapRel=gapRel,
+                logPath=logPath,
+                warmStart=warmStart,
+            )
+
+            from cuopt.linear_programming import data_model  # type: ignore[import-not-found, import-untyped, unused-ignore]
+
+            self.model = data_model.DataModel()
+            self.var_list = None
+            self.solver_params = solverParams
+
+        def findSolutionValues(self, lp, solution):
+            solutionStatus = solution.get_termination_status()
+            if self.msg:
+                print("CUOPT status=", solution.get_termination_reason())
+
+            CuoptStatus = {
+                0: LpStatusNotSolved,  # No Termination
+                1: LpStatusOptimal,  # Optimal
+                2: LpStatusInfeasible,  # Infeasible
+                3: LpStatusUnbounded,  # Unbounded
+                4: LpStatusNotSolved,  # Iteration Limit
+                5: LpStatusNotSolved,  # Timelimit
+                6: LpStatusNotSolved,  # Numerical Error
+                7: LpStatusNotSolved,  # Primal Feasible
+                8: LpStatusNotSolved,  # Feasible Found
+                9: LpStatusNotSolved,  # Concurrent Limit
+            }
+
+            lp.resolveOK = True
+            for var in lp._variables:
+                var.isModified = False
+
+            status = CuoptStatus.get(solutionStatus, LpStatusUndefined)
+            lp.assignStatus(status)
+
+            values = solution.get_primal_solution()
+
+            for var, value in zip(lp._variables, values):
+                var.varValue = value
+
+            if not solution.get_problem_category():
+                # TODO: Compute Slack
+
+                redcosts = solution.get_reduced_cost()
+                for var, value in zip(lp._variables, redcosts):
+                    var.dj = value
+
+                duals = solution.get_dual_solution()
+                for constr, value in zip(lp.constraints.values(), duals):
+                    constr.pi = value
+
+            return status
+
+        def available(self):
+            """True if the solver is available"""
+            return True
+
+        def callSolver(self, lp, callback=None):
+            """Solves the problem with CUOPT"""
+            from cuopt.linear_programming import solver_settings, solver  # type: ignore[import-not-found, import-untyped, unused-ignore]
+
+            self.solveTime = -clock()
+            # TODO: Add callback
+            log_file = self.optionsDict.get("logPath") or ""
+
+            settings = solver_settings.SolverSettings()
+            settings.set_parameter("infeasibility_detection", True)
+            settings.set_parameter("log_to_console", self.msg)
+            if self.timeLimit:
+                settings.set_parameter("time_limit", self.timeLimit)
+            for key, value in self.solver_params.items():
+                if key == "optimality_tolerance":
+                    settings.set_optimality_tolerance(value)
+            gapRel = self.optionsDict.get("gapRel")
+            if gapRel:
+                settings.set_parameter("relative_gap_tolerance", gapRel)
+
+            solution = solver.Solve(lp.solverModel, settings, log_file)
+
+            self.solveTime += clock()
+            return solution
+
+        def buildSolverModel(self, lp):
+            """
+            Takes the pulp lp model and translates it into a COPT model
+            """
+            lp.solverModel = self.model
+
+            if lp.sense == LpMaximize:
+                lp.solverModel.set_maximize(True)
+
+            var_lb, var_ub, var_type, var_name = [], [], [], []
+            obj_coeff = []
+            var_dict = {}
+
+            for i, var in enumerate(lp.variables()):
+                obj_coeff.append(lp.objective.get(var, 0.0))
+                lowBound = var.lowBound
+                if lowBound is None:
+                    lowBound = -np.inf
+                upBound = var.upBound
+                if upBound is None:
+                    upBound = np.inf
+                varType = "C"
+                if var.cat == LpInteger and self.mip:
+                    varType = "I"
+                if var.cat == LpBinary and self.mip:
+                    varType = "I"
+                    lowBound = 0
+                    upBound = 1
+                var_lb.append(lowBound)
+                var_ub.append(upBound)
+                var_type.append(varType)
+                var_name.append(var.name)
+                var_dict[var.name] = i
+                var.solverVar = {
+                    var.name: {"lb": var_lb, "ub": var_ub, "type": var_type}
+                }
+            lp.solverModel.set_variable_lower_bounds(np.array(var_lb))
+            lp.solverModel.set_variable_upper_bounds(np.array(var_ub))
+            lp.solverModel.set_variable_types(np.array(var_type))
+            lp.solverModel.set_variable_names(np.array(var_name))
+
+            rhs, sense = [], []
+            matrix_data, matrix_indices, matrix_indptr = [], [], [0]
+
+            for name, constraint in lp.constraints.items():
+                row_coeffs = []
+                matrix_data.extend(list(constraint.values()))
+                matrix_indices.extend([var_dict[v.name] for v in constraint.keys()])
+                matrix_indptr.append(len(matrix_data))
+                try:
+                    c_sense = sense_conv[constraint.sense]
+                except:
+                    raise PulpSolverError("Detected an invalid constraint type")
+                rhs.append(-constraint.constant)
+                sense.append(c_sense)
+            lp.solverModel.set_csr_constraint_matrix(
+                np.array(matrix_data), np.array(matrix_indices), np.array(matrix_indptr)
+            )
+            lp.solverModel.set_constraint_bounds(np.array(rhs))
+            lp.solverModel.set_row_types(np.array(sense))
+
+            lp.solverModel.set_objective_coefficients(np.array(obj_coeff))
+
+        def actualSolve(self, lp, callback=None):
+            """
+            Solve a well formulated lp problem
+
+            creates a COPT model, variables and constraints and attaches
+            them to the lp model which it then solves
+            """
+            self.buildSolverModel(lp)
+            solution = self.callSolver(lp, callback=callback)
+
+            solutionStatus = self.findSolutionValues(lp, solution)
+            for var in lp._variables:
+                var.modified = False
+            for constraint in lp.constraints.values():
+                constraint.modified = False
+            return solutionStatus
+
+        def actualResolve(self, lp, callback=None):
+            """
+            Solve a well formulated lp problem
+
+            uses the old solver and modifies the rhs of the modified constraints
+            """
+            rhs = lp.solverModel.get_constraint_bounds()
+            sense = lp.solverModel.get_row_types()
+
+            for i, name, constraint in enumerate(lp.constraints.items()):
+                if constraint.modified:
+                    sense[i] = sense_conv[constraint.sense]
+                    rhs[i] = -constraint.constant
+                    constraint.solverConstraint[name]["bound"] = rhs[i]
+                    constraint.solverConstraint[name]["sense"] = sense[i]
+            lp.solverModel.set_constraint_bounds(rhs)
+            lp.solverModel.set_row_types(sense)
+
+            self.callSolver(lp, callback=callback)
+
+            solutionStatus = self.findSolutionValues(lp)
+            for var in lp._variables:
+                var.modified = False
+            for constraint in lp.constraints.values():
+                constraint.modified = False
+            return solutionStatus
diff --git a/pulp/tests/test_pulp.py b/pulp/tests/test_pulp.py
@@ -100,7 +100,11 @@ class PuLPTest(unittest.TestCase):
         solveInst: Optional[Type[LpSolver]] = None
 
         def setUp(self):
-            self.solver = self.solveInst(msg=False)
+            if self.solveInst == CUOPT:
+                # cuOpt requires a user provided time limit for MIP problems
+                self.solver = self.solveInst(msg=False, timeLimit=120)
+            else:
+                self.solver = self.solveInst(msg=False)
             if not self.solver.available():
                 self.skipTest(f"solver {self.solveInst.name} not available")
 
@@ -2233,6 +2237,10 @@ class COPTTest(BaseSolverTest.PuLPTest):
     solveInst = COPT
 
 
+class CUOPTTest(BaseSolverTest.PuLPTest):
+    solveInst = CUOPT
+
+
 class SASTest:
 
     def test_sas_with_option(self):
