Issues

• cuSOLVER potrf is not CUDA-graph capturable on CUDA.jl 6.5: no graph replay with vendor dense fronts (T34 step 6, PLAN §2.5)

  found-by-agentperformanceupstream
  exanauts/SparseDirectSolver.jl#174 · Filter by author michel2323 opened on Oct 30, 2026
  1

• Metal extension follow-ups (need Apple hardware): ComplexF32, MPS command buffers, batched stats reduction

  bugenhancementmetalon-holdperformance
  exanauts/SparseDirectSolver.jl#168 · Filter by author michel2323 opened on Oct 29, 2026 · M2 — Numeric kernels

• In-front Bunch–Kaufman pivoting for the fused pipeline: state of the closed prototypes #135 and #139

  design-recordenhancementfound-by-agentfusedmadnlp
  exanauts/SparseDirectSolver.jl#167 · Filter by author michel2323 opened on Oct 29, 2026

• Proposal: merge the regime-A/B kernel bodies across Cholesky, LDLᵀ and LU (after T34)

  design-recordenhancementfound-by-agentperformance
  exanauts/SparseDirectSolver.jl#163 · Filter by author michel2323 opened on Oct 29, 2026

• Options accepted and silently ignored: max_lu_nnz, host_nthreads, use_cuda_register_memory, schedule = "subtree+level"

  apifound-by-agent
  exanauts/SparseDirectSolver.jl#159 · Filter by author michel2323 opened on Oct 29, 2026

• T34 — fused pipeline: close the sm_89 gap

  found-by-agentfusedperformancetask
  exanauts/SparseDirectSolver.jl#156 · Filter by author michel2323 opened on Oct 29, 2026 · M11 — Performance

• oneAPI.jl shims: uncaught oneMKL exceptions, batched trsm pointer types, host sync per oneMKL call

  found-by-agentoneapiupstream
  exanauts/SparseDirectSolver.jl#148 · Filter by author michel2323 opened on Oct 27, 2026
  1

• T32 — Allocation-free KA dense fallbacks, :auto prefers :ka (issues #112, #53)

  task
  exanauts/SparseDirectSolver.jl#146 · Filter by author michel2323 opened on Oct 27, 2026 · M2 — Numeric kernels

• CUDA.jl 6.5 exports update!, ambiguous with SparseDirectSolver.update! under using CUDA

  apidocumentationfound-by-agentmadnlp
  exanauts/SparseDirectSolver.jl#144 · Filter by author michel2323 opened on Oct 27, 2026

• _tri_decode: integer-search isqrt in the regime-A/B inner loops not timed on CUDA/AMDGPU

  benchfound-by-agentperformance
  exanauts/SparseDirectSolver.jl#142 · Filter by author michel2323 opened on Oct 27, 2026

• T30 — ND partition-tree export/import and ordering cache

  task
  exanauts/SparseDirectSolver.jl#117 · Filter by author michel2323 opened on Oct 9, 2026 · M1 — Symbolic engine

• T29 — Non-uniform batch (BatchedDirectSolver)

  task
  exanauts/SparseDirectSolver.jl#116 · Filter by author michel2323 opened on Oct 9, 2026 · M10 — Non-uniform batch

• T28 — Robustness extras

  task
  exanauts/SparseDirectSolver.jl#115 · Filter by author michel2323 opened on Oct 9, 2026 · M13 — Robustness extras

• Regime-C Cholesky reportedly host-synchronizes per front through its info check (blocks stream overlap of wide fronts); verify

  found-by-agentperformancetriaged
  exanauts/SparseDirectSolver.jl#110 · Filter by author michel2323 opened on Oct 9, 2026 · M11 — Performance
  1

• Update-stack placement assumes level-synchronous execution: out-of-level-order kernels (experiment 5) need private contribution blocks or DAG lifetimes

  found-by-agentperformancetriaged
  exanauts/SparseDirectSolver.jl#109 · Filter by author michel2323 opened on Oct 9, 2026 · M11 — Performance
  1

• Compile time: fewer kernel specialisations per (T, INT)

  ciperformancetriaged
  exanauts/SparseDirectSolver.jl#91 · Filter by author michel2323 opened on Oct 6, 2026
  1

• KKT refactorization and solve are level-bound after the regime-A split: 54–59 launches per refactorization, 62–85 per solve, regime-B fronts on one or two blocks

  found-by-agentperformancetriaged
  exanauts/SparseDirectSolver.jl#82 · Filter by author michel2323 opened on Oct 5, 2026 · M11 — Performance

• Dense :ka/:generic paths and the KA CPU backend allocate in the numeric phase (Julia 1.10, coverage)

  found-by-agenttriaged
  exanauts/SparseDirectSolver.jl#53 · Filter by author claude opened on Oct 2, 2026 · M2 — Numeric kernels
  2

• External — MadNLPGPU integration (in the MadNLP repository)

  externaltask
  exanauts/SparseDirectSolver.jl#28 · Filter by author michel2323 opened on Oct 1, 2026 · M4 — Symmetric indefinite LDLᵀ/LDLᴴ (MadNLP-ready)

• T27 — Hybrid memory / hybrid execute

  task
  exanauts/SparseDirectSolver.jl#27 · Filter by author michel2323 opened on Oct 1, 2026 · M12 — Hybrid modes

• T26 — Segmented fused factorization with dependency counters (issues #82, #109, #110)

  task
  exanauts/SparseDirectSolver.jl#26 · Filter by author michel2323 opened on Oct 1, 2026 · M11 — Performance

• T25 — Split factorization: tiled SYRK, chunked TRSM, split regime-C fronts

  task
  exanauts/SparseDirectSolver.jl#25 · Filter by author michel2323 opened on Oct 1, 2026 · M11 — Performance

• T24 — Partitioned-inverse fused solve (solve_alg = "algo1", issue #82 solve half)

  on-holdtask
  exanauts/SparseDirectSolver.jl#24 · Filter by author michel2323 opened on Oct 1, 2026 · M11 — Performance
  2