MethodError: no method matching QuadraticModels.QPData during JuMP.optimize!

[ChetanAnilMore]

Hello,

I am encountering a MethodError when calling JuMP.optimize! on a model using MadIPM with the CUDSSSolver. It seems there is a version incompatibility between MadIPM's MOI wrapper and QuadraticModels.

Minimal Reproducible Example

Here is the script that triggers the error:

using ExaModels
using CUDA
using KernelAbstractions
using MadNLP
using MadIPM
using CUDSS
using MadNLPGPU
using JuMP

N = 1000
u = rand(Float64, N) .* 0.0 # simplified for MRE
u_gpu = CuArray(u)

model = Model(MadIPM.Optimizer)
set_optimizer_attribute(model, "array_type", CuVector{Float64})
set_optimizer_attribute(model, "linear_solver", MadNLPGPU.CUDSSSolver)

@variable(model, x[1:N])
set_start_value.(x, fill(0.0, N))

@objective(model, Min, 0.5 * sum((x .- 1.0) .^ 2))

for j in 1:N
    @constraint(model, (u[j] * sum(x) - 500.0) .<= 0.0)
end

JuMP.optimize!(model)

ERROR: LoadError: MethodError: no method matching QuadraticModels.QPData(::Float64, ::Vector{Float64}, ::Vector{Float64}, ::SparseMatricesCOO.SparseMatrixCOO{Float64, Int64}, ::SparseMatricesCOO.SparseMatrixCOO{Float64, Int64})

Closest candidates are:
  QuadraticModels.QPData(::T, ::S, ::S, ::M1, ::M2, ::Bool, ::I, ::T) where {T, S, M1<:Union{AbstractMatrix{T}, LinearOperators.AbstractLinearOperator{T}}, M2<:Union{AbstractMatrix{T}, LinearOperators.AbstractLinearOperator{T}}, I<:(AbstractVector{<:Integer})}
   @ QuadraticModels C:\Users\Chetan\.julia\packages\QuadraticModels\u3FxY\src\qpmodel.jl:10

Version Information: 
[4076af6c] JuMP v1.30.1
[4406d55e] MadIPM v0.2.0
[f468eda6] QuadraticModels v0.9.16

[klamike]

Hi @ChetanAnilMore, thanks for opening the issue. We are actually currently in the process of addressing this, waiting on registration for another package before we can do a new release. In the meantime, you can try pinning QuadraticModels to v0.9.14 as they changed the QPData struct in 0.9.15.

[klamike]

There is another issue due to the latest MadNLP releases 🙃

So, to get things working with the current registry, you should either:

• Option A: pin MadNLP to v0.9.1 and QuadraticModels to v0.9.14, or
• Option B: add the following snippet after your using block:

MadIPM.QuadraticModels.QPData(c0, c, v, H, A; regularize = false, selected = 1:length(c), σ = zero(c0)) = MadIPM.QuadraticModels.QPData(c0, c, v, H, A, regularize, selected, σ)

for (k, attribute) in enumerate(fieldnames(MadIPM.MPCSolver))
    fname = "get_$(attribute)"
    sf = Symbol(fname)
    if isdefined(MadNLP, sf)
        @eval begin
            @inline function MadNLP.$(sf)(solver::MadIPM.MPCSolver)
                return getfield(solver, $k)
            end
        end
    else
        @eval begin
            @inline function $(sf)(solver::MadIPM.MPCSolver)
                return getfield(solver, $k)
            end
        end
    end
end

Sorry for the hassle, reshuffling a bunch of stuff right now 😅

[amontoison]

cc @MaxenceGollier

Maxence, I was not aware but you broke MadIPM.jl with the last releases of QuadraticModels.jl 🫠

To the best of my knowledge QuadraticModels.QPData is not exported, so it was difficult to know that we were using it here.

Nobody added new features in QuadraticModels.jl since years (end of the development of RipQP.jl ).

[amontoison]

@ChetanAnilMore I created a branch backport-0.2, it should fix your issue.
You can install it with pkg> add MadIPM#backport-0.2.

Update: I did a new release 0.2.1.

[amontoison]

@frapac @klamike @apozharski

When this kind of thing happens, don't hesitate to make an "out-of-tree" release.
I just released v0.2.1.
It takes 5 minutes, we don't want to leave users with broken releases for weeks...

[MaxenceGollier]

Hi @amontoison.

cc @MaxenceGollier

Maxence, I was not aware but you broke MadIPM.jl with the last releases of QuadraticModels.jl 🫠

To the best of my knowledge QuadraticModels.QPData is not exported, so it was difficult to know that we were using it here.

Nobody added new features in QuadraticModels.jl since years (end of the development of RipQP.jl ).

I apologize, I should have released 0.10 instead of 0.9.15.
As I understand, the issue is that you contruct QPData without using the user interface

@inline QPData(c0, c, H, A) = ... 

which has been correctly updated in the release of 0.9.15. Is there a reason for this ?
Using the above is safer and prevents this kind of issue.

@amontoison, with my recent update of the breakage workflow in JSO, we could easily add in addition to the JSO deps, the MadNLP deps to the workflow in order to avoid this in the future, what do you think ?

cc. @dpo.

[apozharski]

I am going to push back a bit on the whole breakage workflow (also relevant for madsuite-org/MadNLP.jl#625 @frapac). I don't think having every package test whether it breaks its dependencies is sustainable moving forward (and frankly is a waste of compute resources). We should instead more aggressively move towards defining public APIs in a v1.0.0 (as is the semver way https://semver.org/), especially in such low level packages such as NLPModels, QuadraticModels, et cetera. I understand the desire to not prematurely calcify the API but I think this leads to both the eternal dependency hell we find ourselves in, and the constant "just one more feature" API creep.

If you would be amenable to it @dpo I would be happy to some time in the coming weeks write up a brief design document for a v1.0.0 NLPModels ecosystem (as we briefly discussed in Edinburgh).