Schur design constraints and sparse Schur complement solver - #623

Open

michel2323 wants to merge 31 commits into

masterfrom

schur-design-constraints

Commits on Jul 20, 2026

[MadNLPGPU] cuDSS: recover inertia under matching; guard matching off ubatch
michel2323
committed
0fcad01
[MadNLPGPU] Sparse Schur complement solver with design constraints
michel2323
committed
1ae8dfa
[MadNLPGPU] Default SchurComplementKKTSystem to RelaxEquality; bump cuDSS IR 1->5
michel2323
committed
08fb504
Set CUDSS IR=0 for Schur KKT tests
michel2323
committed
0361f0f
CUDSS Schur KKT error handling
michel2323
committed
a81af18
[MadNLPGPU] cuDSS 0.8: disarm the IR_FAILED gate (cudss_ir_tol = 0.0)
michel2323
committed
fe0c414
[MadNLPGPU] Deterministic Schur assembly: reproducible GPU two-stage convergence
michel2323
committed
ed7a296
[MadNLPGPU] Remove stale cuDSS IR-failure Schur test
michel2323
committed
107509e
bound_relax_factor: relax condensed (Sparse+Schur) KKT by tol, consistently on CPU/CUDA/ROCm
michel2323
committed
4cf1acc
[KKT] Rename SchurComplementKKTSystem -> SchurComplementCondensedKKTSystem
michel2323
committed
2f829c2
[MadNLPGPU] Fix GPU two-stage Schur refinement mul!: correct full-symmetric Hessian
michel2323
committed
79678d9
[MadNLPGPU] Work around broken cuDSS multi-RHS uniform-batch solve in Schur build
michel2323
committed
d519548
[KKT] Schur is not is_kkt_dense: reject dense BFGS with a clear options error
michel2323
committed
0d428a8
[options] Extract CondensedKKTSystems union (dedupe 6x copy-paste)
michel2323
committed
6aac1ba
[KKT] Deprecate SchurComplementKKTSystem alias; clearer EnforceEquality error
michel2323
committed
203a934
[MadNLPGPU] Error clearly for SchurComplementCondensedKKTSystem on ROCm
michel2323
committed
0a62232
[KKT] Unify CPU/GPU Schur create_kkt_system kwarg surface
michel2323
committed
1825df6
[KKT] CPU Schur: reduce elimination + coupling buffers to the m coupled columns
michel2323
committed
74d7dff
[test] Cover the m < nd (uncoupled design variables) Schur path
michel2323
committed
d57195e
[docs] Schur complement is sparse (lower-triangular CSC), not dense
michel2323
committed
12ff1bb
[MadNLPGPU] Revert orphaned LapackCUDASolver Bunch-Kaufman inertia override
michel2323
committed
0ed713c
[MadNLPGPU] Restore global cudss_ir_tol default to 1e-8 (scope 0.0 to Schur)
michel2323
committed
1a4e63f
[MadNLPGPU] Schur compress_jacobian!: SUM duplicate Jacobian COO entries
michel2323
committed
a282c30
[MadNLPGPU] Schur GPU: warn when user linear_solver is discarded
michel2323
committed
af4e996
[MadNLPGPU] cuDSS: log status before erasing it; drop redundant Schur double-wrap
michel2323
committed
3f3b9d1
[MadNLPGPU] Remove 8 dead Schur assembly kernels
michel2323
committed
f681d2b
[KKT] Collapse the dead nd_aug alias to nd
michel2323
committed
e8518b3
[MadNLPGPU] Remove dead cuDSS multi-RHS Schur descriptors
michel2323
committed
942556e
[MadNLPGPU] Batch the Schur reduction/back-substitution GEMMs
michel2323
committed
84fd90d
[MadNLPGPU] Schur det_cache: type-stable hot paths via function barriers
michel2323
committed
6dd1a69

Commits on Jul 28, 2026

[KKT] Include the per-scenario blocks in the Schur inertia test
Github action runner
committed
dd0d099

Navigation Menu

Uh oh!

Schur design constraints and sparse Schur complement solver - #623

Schur design constraints and sparse Schur complement solver#623
michel2323 wants to merge 31 commits into
masterfrom
schur-design-constraints

Commits on Jul 20, 2026

[MadNLPGPU] cuDSS: recover inertia under matching; guard matching off ubatch

[MadNLPGPU] Sparse Schur complement solver with design constraints

[MadNLPGPU] Default SchurComplementKKTSystem to RelaxEquality; bump cuDSS IR 1->5

Set CUDSS IR=0 for Schur KKT tests

CUDSS Schur KKT error handling

[MadNLPGPU] cuDSS 0.8: disarm the IR_FAILED gate (cudss_ir_tol = 0.0)

[MadNLPGPU] Deterministic Schur assembly: reproducible GPU two-stage convergence

[MadNLPGPU] Remove stale cuDSS IR-failure Schur test

bound_relax_factor: relax condensed (Sparse+Schur) KKT by tol, consistently on CPU/CUDA/ROCm

[KKT] Rename SchurComplementKKTSystem -> SchurComplementCondensedKKTSystem

[MadNLPGPU] Fix GPU two-stage Schur refinement mul!: correct full-symmetric Hessian

[MadNLPGPU] Work around broken cuDSS multi-RHS uniform-batch solve in Schur build

[KKT] Schur is not is_kkt_dense: reject dense BFGS with a clear options error

[options] Extract CondensedKKTSystems union (dedupe 6x copy-paste)

[KKT] Deprecate SchurComplementKKTSystem alias; clearer EnforceEquality error

[MadNLPGPU] Error clearly for SchurComplementCondensedKKTSystem on ROCm

[KKT] Unify CPU/GPU Schur create_kkt_system kwarg surface

[KKT] CPU Schur: reduce elimination + coupling buffers to the m coupled columns

[test] Cover the m < nd (uncoupled design variables) Schur path

[docs] Schur complement is sparse (lower-triangular CSC), not dense

[MadNLPGPU] Revert orphaned LapackCUDASolver Bunch-Kaufman inertia override

[MadNLPGPU] Restore global cudss_ir_tol default to 1e-8 (scope 0.0 to Schur)

[MadNLPGPU] Schur compress_jacobian!: SUM duplicate Jacobian COO entries

[MadNLPGPU] Schur GPU: warn when user linear_solver is discarded

[MadNLPGPU] cuDSS: log status before erasing it; drop redundant Schur double-wrap

[MadNLPGPU] Remove 8 dead Schur assembly kernels

[KKT] Collapse the dead nd_aug alias to nd

[MadNLPGPU] Remove dead cuDSS multi-RHS Schur descriptors

[MadNLPGPU] Batch the Schur reduction/back-substitution GEMMs

[MadNLPGPU] Schur det_cache: type-stable hot paths via function barriers

Commits on Jul 28, 2026

[KKT] Include the per-scenario blocks in the Schur inertia test