François Pacaud

Add support to dense Jacobian and Hessian in API

18

Together with @sshin23, we are working on rebasing the interface of MadNLP on NLPModels.jl (see https://github.com/sshin23/MadNLP.jl/pull/69). In that context, we are wondering whether it would be possible to add new...

Add methods for dense Hessian and dense Jacobian

2

Continuation of #376

A new API is available for CUSPARSE, deprecating part of the functions already wrapped in CUSPARSE.jl. #### Question Can we adapt CUSPARSE's API to allow the user to store the...

Add support for equality constraints in DenseCondensedKKTSystem

1

Right now, the option `DENSE_CONDENSED_KKT_SYSTEM` does not reduce the equality constraints and focuses only on the inequality constraints. One has to use an elastic relaxation to reformulate the equality constraints...

Time-to-first-solve (TTFS)

2

Now that the `StridedOneVector` PR #174 has been merged, we can focus on the time to first solve. Using the following script: ```julia using MadNLP using MadNLPTests nlp = MadNLPTests.HS15Model()...

Like the line-search and the barrier update, the primal-dual regularization should be modular too, to let the choice to the user to extend the regularization algorithm.

We should allow the user to change the update strategy for the barrier term `mu` (fixed, monotone, adaptive, Mehrotra predictor-corrector type).

At the moment MadNLP is using a classical backtracking line-search. We should implement a modular way to specify the line-search algorithm, and give the possibility to the user to use...

See: https://github.com/sshin23/MadNLP.jl/pull/73#discussion_r705832963 This would allows to run MadNLP on the GPU with inertia based regularization. The main showstopper is the sequential nature of the inertia computation algorithm: we may want...