Integration with cuopt (open source)

[sharabhshukla]

What language and solver does this apply to?
All languages supported by OR-Tools (Python, C++, Java, C#)
Primarily: Linear Solver (LP/MILP interface)
Potentially: CP-SAT (future exploration)

Describe the problem you are trying to solve.
I would like to leverage GPU-accelerated LP and MILP solving capabilities provided by cuOpt, which is now open source, within the OR-Tools ecosystem.
Currently, OR-Tools supports several LP/MILP backends such as GLOP, SCIP, CBC, Gurobi, and CPLEX through the MPSolver interface. However, there is no native GPU-accelerated backend available.
For large-scale linear programs and mixed-integer linear programs (e.g., supply chain optimization, network flow, energy optimization, scheduling), CPU-based solvers become a bottleneck. cuOpt provides GPU-native optimization algorithms that can significantly accelerate these workloads. The goal is to integrate cuOpt as an optional backend solver in OR-Tools' Linear Solver interface.

Describe the solution you'd like
I propose adding cuOpt as a new backend to OR-Tools' Linear Solver (MPSolver) framework.
Specifically:
Implement a new solver type:
MPSolver::CUOPT_LINEAR_PROGRAMMING
MPSolver::CUOPT_MIXED_INTEGER_PROGRAMMING
Create a wrapper class similar to existing backends (e.g., GurobiInterface, ScipInterface) that:
Translates OR-Tools MPModelProto into cuOpt model format
Transfers model data to GPU
Invokes cuOpt solve routines
Retrieves solution values and status
Maps cuOpt statuses to OR-Tools statuses
Expose cuOpt through:
Python API (pywraplp)
C++ API
Java and C# wrappers (via existing OR-Tools binding pipeline)
Support:
LP models
MILP models
Warm starts (if supported by cuOpt)
Basic parameter configuration (tolerances, time limits, threads, GPU selection)
Optional:
Automatic GPU detection
Fallback to CPU solver if no GPU is available
This would allow users to simply switch solvers:
from ortools.linear_solver import pywraplp
solver = pywraplp.Solver.CreateSolver("CUOPT_MIXED_INTEGER_PROGRAMMING")
without changing modeling code.

Describe alternatives you've considered
Well today one can use pdlp's to solve large LP's but there is no direct way I know where I can run my ortools models written in c++ to use any GPU based solvers, having that option in or-tools only makes or-tools richer in terms of solver eco system and cuopt is open source so that presenta a fantastic opportunity

Additional context
More details about cuopt can be found on this link including the source code:
https://github.com/NVIDIA/cuopt

[lperron]

we will not integrate cuopt. Too much work, too many limitations.

cu-pdlp should be integrated in highs

Note: pywraplp, MPsolver (same thing) is in maintenance mode. Work should go into ModelBuilder (small and easy) or MathOpt (recommended)

[lperron]

Let me clarify, PR welcome. But it will be purely community supported.