tril_to_full! function in QR and LU algorithms error on GPU

[dlcole3]

Currently, I get an error when trying to run with MadNLPLapackGPU.QR or MadNLPLapackGPU.LU. The tril_to_full!() function is only defined for objects of type Matrix, so the function doesn't work with CuArray types on the GPU. An example script that results in this error is shown bleow.

using MadNLP, ADNLPModels, MadNLPGPU, CUDA, NLPModels

function MadNLP.jac_dense!(nls, x, jac) 
    NLPModels.increment!(nls, :neval_jac)
    copyto!(jac, zeros(size(jac)))
end

function MadNLP.hess_dense!(nls, x, w1l, hess; obj_weight = 1.0)
    NLPModels.increment!(nls, :neval_hess)
    copyto!(hess, [0 0; -20 0])
end

F(x) = [x[1] - 1.0; 10 * (x[2] - x[1]^2)]
x0 = [-1.2; 1.0]
nls = ADNLSModel(F, x0, 2)

madnlp_options = Dict{Symbol, Any}(
:kkt_system=>MadNLP.DENSE_CONDENSED_KKT_SYSTEM,
:linear_solver=>MadNLPLapackGPU,
:print_level=>MadNLP.DEBUG,
)

linear_solver_options = Dict{Symbol, Any}(
:lapackgpu_algorithm => MadNLPLapackGPU.QR,
)

TKKTGPU = MadNLP.DenseCondensedKKTSystem{Float64, CuVector{Float64}, CuMatrix{Float64}}
opt = MadNLP.Options(; madnlp_options...)
ips = MadNLP.InteriorPointSolver{TKKTGPU}(nls, opt; option_linear_solver=copy(linear_solver_options))
MadNLP.optimize!(ips)

[sshin23]

even madnlp(nls) doesn't work, and it seems that the issue is caused by

get_nnzj(nls) != get_nnzj(nls.meta)

@dpo @abelsiqueira is this an expected behavior? If a solver wants to solve NLS problem as a generic NLP, does the solver need to query nnzj and nnzh from nls.meta, not nls?

[dpo]

@sshin23 get_nnzj(nls) is the same as get_nnzj(nls.nls_meta), which should return the number of nonzeros in the Jacobian of the residual:
https://github.com/JuliaSmoothOptimizers/NLPModels.jl/blob/main/src/nls/tools.jl#L11

By contrast, get_nnzj(nls.meta) would return the number of nonzeros in the Jacobian of the constraints (not counting bound constraints), if your NLSModel had any. In addition, nnzj was recently split into lin_nnzj and nln_nnzj (linear/nonlinear constraints).

[sshin23]

Thanks, @dpo

Then MadNLP should always query nnz information from nlp.meta. @dlcole3 feel free to create a PR for the fix and adding this to MadNLPTest