Add option to print the solvers in memory representation

[joaquimg]

Writing?

model = Model(HiGHS.Optimizer)
@variable(model, p1, set = Parameter(1.0))
@variable(model, p2 == 1.0)
write_to_file(model, "model.mps")
optimize!(model)
highs = unsafe_backend(model).inner
tmp_filename = "model.mps"
HiGHS.Highs_writeModel(highs, tmp_filename)

is hard for non-super-heavy users.

This is helpful for people with minimal understanding of solution methods.

[odow]

is hard for non-super-heavy users.

Should these users be printing out the internal model? I wouldn't really be in favor of them doing that.

This is also quite solver-dependent. Maybe it just needs better documentation in each solver's README.

Having said that, perhaps we could make MOI.write_to_file(::HiGHS.Optimizer, filename::String) work, and have write_to_file(model, "model.mps"; use_internal_solver_method = true)

[odow]

I looked into this. The current isn't too bad:

using JuMP, HiGHS
model = Model(HiGHS.Optimizer)
@variable(model, x >= 1)
optimize!(model)
Highs_writeModel(unsafe_backend(model), "model.mps")

we could perhaps make it

using JuMP, HiGHS
model = Model(HiGHS.Optimizer)
@variable(model, x >= 1)
optimize!(model)
MOI.write_to_file(unsafe_backend(model), "model.mps")

I'm not really sure who this method is aimed at. Someone who wants to deeply introspect the solver's model, knowing that there is a bug somewhere between JuMP and the solver, so JuMP.write_to_file(model) doesn't work, but also has no understanding of Highs_writeModel?

[blegat]

It's also sometimes useful when you're transitioning from ... to JuMP and get different answers and want to check what's the difference. MOI.write_to_file(unsafe_backend(model), "model.mps") is the right way to do it, I don't think we need other changes

[joaquimg]

I like:
MOI.write_to_file(unsafe_backend(model), "model.mps")
Documenting it should solve the issue.
Some places where is might be useful:
1 - debug tutorial
2 - transitioning tutorial
3 - reference in the write_to_file docstring
4 - the solver readme

[odow]

Came up https://discourse.julialang.org/t/using-second-order-cone-constraints-with-logarithms-in-jump-and-gurobis-c-api/125769/7?u=odow

[odow]

I think there's an argument for (3) and (4). Strong disagreement about (2), and moderate disagreement about (1).

In almost every case, write_to_file(model, filename) should be sufficient. If someone is using Gurobi AND they are using things like nonlinear AND they want Gurobi's custom MPS format to send support, then there is an argument. But I don't see how an MPS file is useful to someone debugging unless there is a bug in the MOI file writers.

[odow]

Re (4): I opened a PR do document how to write an MPS file for support in Gurobi: jump-dev/Gurobi.jl#613

My other issue with something generic is that it really applies only to Gurobi/CPLEX/Xpress/HiGHS/Mosek. It won't apply for Ipopt, or SCS, or any number of other solvers.