v0.10.0

MadNLP v0.10.0

Diff since v0.9.2

Highlights: the main solve! loop is now allocation-free in steady state — ~99% of per-iteration allocations have been removed by replacing mapreduce-based reductions with explicit loops and extracting backend-specific kernels so that GPU code paths can plug in without touching the core. Plus a CUDA.jl v6 / CUDSS 0.7 / Julia 1.12 update for MadNLPGPU.

Performance

• Remove ~99% of per-iteration allocations from solve!. get_varphi, get_inf_du, get_inf_compl, get_min_complementarity, and friends in src/IPM/kernels.jl are rewritten as explicit loops instead of mapreduce. count_lu_bounds replaces the boolean-mask + sum pattern in get_vars_info / get_cons_info. Indexed updates in _set_aug_diagonal! are wrapped in @views to avoid temporary copies. The unused dense jac_buffer allocation in the NLPModels callback is removed. (#603)
• New populate_RR_nn! kernel for restoration-phase initialization, factored out so backends can override it. (#603)

Breaking changes

• RR field on MadNLPSolver is now typed Union{Nothing,RobustRestorer{T,VT}} (was RobustRestorer{T}). Downstream code that holds a typed reference to RobustRestorer needs the extra type parameter. (#603)
• get_vars_info and get_cons_info are now restricted to AbstractMadNLPSolver. (#603)
• New extension points (_get_varphi, get_varphi, get_inf_du, get_inf_compl, get_min_complementarity, get_varphi_d, get_alpha_max, get_alpha_z, restoration-phase get_*_R variants, populate_RR_nn!, count_lu_bounds, SubVector) are exported from MadNLP for backends to specialize. (#603)

---

MadNLPGPU v0.10.0

Highlights: bumped to CUDA.jl v6 and CUDSS 0.7, and GPU specializations of the new MadNLP core kernels.

Dependencies

• Require CUDA.jl ≥ 6.0.0 and CUDSS ≥ 0.7. (#615)
• Depend on the CUDA subpackages (CUDACore, cuBLAS, cuSOLVER, cuSPARSE) instead of the CUDA meta-package. The CUDA extension now activates from these subpackages directly, so MadNLPGPU works alongside any CUDA.jl v6.x without forcing the meta-package into the project. (#615)
• AMDGPU.jl ≥ 2.

New features

• GPU specializations of the IPM kernels refactored out in MadNLP #603 (get_varphi, get_inf_du, get_inf_compl, get_min_complementarity, restoration populate_RR_nn!, etc.), dispatched on AbstractGPUVectorOrSubVector. The GPU path keeps using mapreduce while the CPU path uses explicit loops — both avoid the per-iteration allocations that the prior code had. (#603)

Fixes

• Work around a Julia 1.12 regression with the global solver state in MadNLPGPU.jl. (#615)

Compatibility

• Requires MadNLP ≥ 0.10.0.

Merged pull requests:

• Removing 99% of allocations in solve! (#603) (@apozharski)
• compat CUDA@6.0.0 CUDSS@0.7 (#615) (@michel2323)