Default behavior for nln_nnzj

[sshin23]

Hi JSO team!

Recent change in #468 is causing errors in ExaModels.jl and MadNLP.jl, both of which only sets nnzj. I was wondering if the default behavior for nln_nnzj be made e.g., nln_nnzj = nnzj - lin_nnzj?

[amontoison]

Hi Sungho!

We will have a similar issue if the user only provides nln_nnzj.
I'm wondering if the best solution is not to fix it in MadNLP.jl and ExaModels.jl like you did.

cc @tmigot

[tmigot]

Even though I agree with @amontoison , it makes sense to me that user not implementing the linear API shouldn't have to specify nln_nnzj. We could update version 0.21.2 with #477 if you both agree.

[sshin23]

@tmigot I agree that user not implementing the linear API shouldn't have to specify nln_nnzj. From the user's perspective, it will be a net improvement. Of course, there's something I'm missing; we will respect NLPModels.jl dev's decision. For now, we can remove NLPModels.jl v0.21 from the compat entry of MadNLP and ExaModels

[tmigot]

@sshin23 released 21.3 should fix your issue

[sshin23]

Thanks, @tmigot and @amontoison