v0.9.2

MadNLP v0.9.2

Diff since v0.9.1

Highlights: new SchurComplementKKTSystem for two-stage stochastic programs (CPU + GPU), an API cleanup so solver and wrapper internals are accessed through getter functions, and a MUMPS 5.8.x fix.

New features

• Add SchurComplementKKTSystem for exploiting the block-arrowhead structure of two-stage stochastic programs. Reduces each Newton step to a dense nd × nd design-variable system via per-scenario eliminations. CPU variant uses Umfpack for the per-scenario blocks by default; GPU variant lives in the MadNLPGPU CUDA extension and uses batched cuDSS + strided-batched CUBLAS. (#598)

API changes

• Solver member access is now through getter functions instead of direct field access. Downstream code that reaches into MadNLPSolver fields will need to migrate. (#497)
• Wrapper models use getters for inner-model access. (#595)
• solve is now re-exported from SolverCore, aligning MadNLP with the broader JuliaSmoothOptimizers ecosystem. (#601)

Fixes

• Update MUMPS bindings for the 5.8.x ABI. (#613)

Documentation

• Fix mu_init in the warm-start tutorial. (#607)

---

MadNLPGPU v0.9.0

Highlights: CUDA/CUDSS and AMDGPU are now weak dependencies, and the new Schur-complement KKT system has a GPU backend.

• CUDA/CUDSS moved to weak deps, GPU backend code deduplicated. Installing MadNLPGPU no longer pulls CUDA on machines without it; the CUDA-specific code (LapackCUDASolver, CUDSSSolver, GPU Schur) lives in the MadNLPGPUCUDAExt extension and loads only when both CUDA.jl and CUDSS.jl are present. The AMDGPU backend follows the same pattern via MadNLPGPUAMDGPUExt. (#596)
• GPU SchurComplementKKTSystem: batched per-scenario factorizations via cuDSS, Schur accumulation via CUBLAS strided-batched GEMM. (#598)
• Requires MadNLP ≥ 0.9.2.

Known issue — V100 / Volta (sm_70):
CUDA Toolkit 13 dropped support for Volta GPUs. Starting at CUDA.jl v5.10 (March 2026), Pkg may resolve to a CUDA.jl that auto-selects the CUDA 13 runtime, which then fails on V100. MadNLPGPU's compat allows this. Until JuliaGPU/CUDA.jl#3134 ships in a tagged release, V100 users should pin to a CUDA 12.x toolkit by running once per environment:

using CUDA
CUDA.set_runtime_version!(v"12.9")    # any 12.x works

Alternatively, pin CUDA = "~5.9" in your project's Project.toml.

---

MadNLPTests v0.6.1

• Add Schur-complement two-stage stochastic test fixtures used by MadNLP and MadNLPGPU Schur tests. (#598)

Merged pull requests:

• Overhaul solver member access via functions (#497) (@apozharski)
• Wrappers should use getters for nlp models meta (#595) (@apozharski)
• Add SchurComplementKKTSystem for two-stage stochastic programs (#598) (@michel2323)
• Revert and use function solve from SolverCore (#601) (@frapac)
• [doc] Fix mu_init in warm-start tutorial (#607) (@frapac)
• [Mumps] Fix the MUMPS structure for v5.8.x (#613) (@frapac)

Closed issues:

• Documentation (#93)
• Restoration failed when using MadNLP for ODE parameter estimation (#264)
• Using weak dependencies (#265)
• lost of precision in constraints using MadNLP.SparseCondensedKKTSystem? (#593)
• Consider not exporting own solve! (#597)
• Performance regression in v0.9.1 compared to v0.8.12 with Mumps (#602)
• Warm-start Tutorial Appears to be Out-of-Date (#606)