A nonlinear programming solver based on the filter line-search interior point method (as in Ipopt) that can handle/exploit diverse classes of data structures, either on host or device memories.
The following example shows how to solve the HS15 problem with JuMP and MadNLP:
using JuMP, MadNLP
model = Model()
@variable(model, x1 <= 0.5)
@variable(model, x2)
@objective(model, Min, 100.0 * (x2 - x1^2)^2 + (1.0 - x1)^2)
@constraint(model, x1 * x2 >= 1.0)
@constraint(model, x1 + x2^2 >= 0.0)
JuMP.set_optimizer(model, MadNLP.Optimizer)
JuMP.set_optimizer_attribute(model, "max_iter", 100)
JuMP.set_optimizer_attribute(model, "print_level", MadNLP.INFO)
optimize!(model)MadNLP can be installed directly via the Julia package manager:
pkg> add MadNLPFor NVIDIA GPUs, install cuMadNLP. rocMadNLP (AMD GPUs), oneMadNLP (Intel GPUs) and mtlMadNLP (Apple GPUs) are not registered yet, see the
installation page. They do
not load MadNLP, so load it as well to solve a model on the CPU or to use JuMP. The HSL
linear solvers are available in MadNLP once a licensed HSL library is installed; the
Pardiso and Krylov.jl solvers become available when Pardiso.jl or Krylov.jl is loaded
alongside MadNLP.
This repository contains the packages below. All of them except MadNLPTests share one
version number and are released together.
| Package | Role |
|---|---|
MadCore |
Callbacks, KKT systems, linear solver interface, options and the helpers shared by the interior-point solvers |
MadCoreGPU |
GPU KKT systems, kernels and the GPU linear solvers, as extensions |
MadNLPCore |
The filter line-search interior-point method |
MadNLP |
CPU entry point, with MUMPS as default linear solver |
cuMadNLP |
NVIDIA entry point, with cuDSS as default linear solver |
rocMadNLP |
AMD entry point, with SparseDirectSolver.jl as default linear solver; not registered |
oneMadNLP |
Intel entry point, with SparseDirectSolver.jl as default linear solver; not registered |
mtlMadNLP |
Apple entry point, with SparseDirectSolver.jl as linear solver; Float32 only; not registered |
MadNLPTests |
Test instances shared by the packages |
MadNLPGPU |
Deprecated; loads cuMadNLP |
If you use MadNLP.jl in your research, we would greatly appreciate your citing it.
@article{pacaud2026condensed,
title = {Condensed interior-point methods for scalable nonlinear programming on GPUs},
author = {Pacaud, Fran{\c{c}}ois and Shin, Sungho and Montoison, Alexis and Schanen, Michel and Anitescu, Mihai},
journal = {Mathematical Programming Computation},
pages = {1--39},
year = {2026},
publisher = {Springer}
}To cite the software itself, use the Zenodo archive 10.5281/zenodo.6335552, which always resolves to the latest release. Each release also has its own version DOI, listed on that page.
- Please report issues and feature requests via the GitHub issue tracker.
- Questions are welcome at GitHub discussion forum.