Merge pull request #74 from harmoniqs/feat/madnlp-gpu-passthroughs

Add MadNLP pass-through fields (linear_solver, array_type, etc.)

Author: jack-champagne

ext/MadNLPSolverExt/solver.jl

@@ -208,7 +208,12 @@ function DirectTrajOpt.set_options!(optimizer::AbstractOptimizer, options::MadNL
         if name in ignored_options
             continue
         end
-        # TODO: allow internal defaults, i.e. do not set the internal options dict unless the user actually specified the associated opt
+        # `nothing` means "use MadNLP's own default" — don't overwrite the optimizer's
+        # internal dict in that case. Applies to the pass-through fields
+        # (linear_solver, array_type, kkt_system, cudss_ordering).
+        if value === nothing
+            continue
+        end
         if name == :print_level
             optimizer.options[name] = MadNLP.LogLevels(value)
         elseif name == :hessian_approximation

src/solvers/madnlp_solver/options.jl

@@ -7,6 +7,13 @@ export MadNLPOptions
     print_level::Int = 3 # (MadNLP.TRACE::MadNLP.LogLevels = 1, ..., MadNLP.ERROR::MadNLP.LogLevels = 6)
     hessian_approximation::String = "exact" # (exact = MadNLP.ExactHessian, compact_lbfgs = MadNLP.CompactLBFGS) # no other QN methods supported in conjunction with MadNLP.SparseCallback
 
+    # Pass-throughs consumed by MadNLP's MOI layer (not by MadNLP itself);
+    # leave as `nothing` to use MadNLP defaults. Only forwarded when non-nothing.
+    linear_solver::Any = nothing  # e.g. MadNLPGPU.CUDSSSolver, MadNLP.LapackCPUSolver
+    array_type::Any = nothing  # e.g. CUDA.CuArray for GPU
+    kkt_system::Any = nothing  # e.g. MadNLP.SparseUnreducedKKTSystem
+    cudss_ordering::Any = nothing  # e.g. MadNLPGPU.AMD_ORDERING
+
     # # Only supported by DirectTrajOpt._solve, as an optional kwarg override of `hessian_approximation`;
     # #   `hessian_approximation = eval_hessian ? "exact" : "compact_lbfgs"`
     # eval_hessian::Bool = true