[0x17]

• Build both workflows against the released cuOpt 26.10 wheels from PyPI: libcuopt-mathopt-cuXX==26.10.* + libcuopt-client==26.10.*
  • 26.10 splits libcuopt into component wheels (libcuopt-mathopt-cuXX, libcuopt-client, libcuopt-routing-cuXX); the link only needs mathopt and client
  • all dependencies resolve from plain PyPI, so no RAPIDS nightly index, --prerelease=allow or pypi.nvidia.com anymore; ==26.10.* without --prerelease only matches final releases (picks up 26.10.x hotfixes)
  • during development the branch was built against 26.10 nightlies (Build against latest cuOpt nightly wheels from RAPIDS nightly index #23); the earlier a223/a224 pin and the revert 4855ada are only history now
  • CONTRIBUTING.md local build instructions updated accordingly (were still cuopt-cu13==26.8.*)
• Adapt link and bundles to the split cuOpt libraries (both workflows and build-link.sh)
  • link -lcuopt_mathopt -lcuopt_client, add -I libcuopt_client/include (status_codes.h ships in the client wheel)
  • bundle libcuopt_mathopt.so + libcuopt_client.so and cuOpt's cuDSS threading layer libcudss_mtlayer_cuopt.so; libcudss_mtlayer_gomp.so is no longer bundled
  • libcuopt_client.so needs system libssl.so.3/libcrypto.so.3/libz.so.1 (not bundled)
• Support multi-GPU PDLP (Add C API support for multiGPU PDLP  NVIDIA/cuopt#1958)
  • optcuopt.def: num_gpus now accepts -1..72, new multigpu_pdlp_partitioner option
  • skip initial primal/dual solutions for multi-GPU PDLP, cuOpt rejects them
  • README section on usage and limitations
  • the cuOptCreateRangedProblem workaround for [BUG] Multi-GPU PDLP sees 0 constraints for problems created with cuOptCreateProblem (row types + RHS) NVIDIA/cuopt#2042 is gone again (Drop cuOptCreateRangedProblem workaround for multi-GPU PDLP #24), fixed upstream in Fix bug #2042 absence of logs on mPDLP error with C API and not using rhs when present NVIDIA/cuopt#2091
• Update optcuopt.def to cuOpt 26.10, checked against the headers of the released v26.10.00 wheels
  • remove 4 options dropped upstream (mip_hyper_heuristic_presolve_time_ratio, mip_hyper_heuristic_presolve_max_time, mip_hyper_diving_min_node_depth, mip_hyper_submip_node_limit_base)
  • concurrent_nnz_cutoff renamed to large_nnz_threshold upstream without an alias (Sync cuOpt options with NVIDIA/cuopt main (8eb20bd) #25), so option files using the old name now get Unknown option
  • new method 4 (primal simplex) and barrier_dual_initial_point 2 (SeDuMi-style), fix changed default of mip_hyper_heuristic_related_vars_time_limit
  • new options, e.g. primal_simplex_pricing, mip_rens, mip_mutation, barrier regularization, Curtis-Reid scaling, postsolve_info, mip_hyper_presolve_indicator_strengthening, degenerate pivots (mip_dual_degenerate_feasibility_pump, mip_primal_degenerate_pivots, mip_dual_degenerate_pivots), pdlp_hyper_reduced_matrix, pdlp_hyper_project_initial_primal
  • left out: sequence_solve (Python re-solve cache only), log_file/log_to_console (set by the link)
• Report iterations and nodes to GAMS via the new solution attributes (cuOptGetSolutionIntAttribute)
  • LP iterations, MIP nodes and simplex iterations; node-limit detection now uses the actual node count
  • guarded by #ifdef, still compiles against 26.08 headers
• Pass GAMS nodlim to cuOpt node_limit (was ignored before)
• Bug fix: drop the PDLP reduced-cost workaround in gmscuopt.c, cuOpt now returns correct reduced costs (Correctly return reduced costs for PDLP (stable3) NVIDIA/cuopt#1797)
• Merged main (Fix example notebooks for current release zips #21, README quickstart, Colab badges, performance notes and community files #22); the only conflict (notebook links in README.md) resolved in favour of main's Colab badges
• Verified
  • released 26.10.0 wheels (v26.10.00, e93f7dc5; mathopt and client from the same commit), CUDA 13, single RTX A1000: link builds without warnings, tests/test-fetch-cuoptlink.py 29/29
  • gamslib test against the released wheels: 73/73 match CPLEX (poutil hits its 300 s limit with the reference objective, run-to-run variance)
  • examples/models/regression_tests/run_tests.sh against the released wheels: 10/10 pass
  • earlier against nightly a285 (cuOpt 74923bbd): gamslib test 73/73 match CPLEX
  • multi-GPU code path (method 1, num_gpus -1, presolve 0, runs even on 1 GPU) with nightly a278: trnsport optimal, gamslib 73/73
  • CI builds and packages CUDA 12 and CUDA 13 on x86_64 and ARM64 from PyPI only (no solves in CI)
  • not tested on a machine with several GPUs
• Not done
  • cuOptSetLogCallback for a live log, since it only receives lines from the calling thread (a MIP log loses ~75% of its lines)
  • example notebooks still show output of the released v0.0.8e link, re-run after the next release

[mlubin]

Huh? please let us know about these types of issues :)
@Bubullzz

[0x17]

Ah, you're right. Didn't have the time to properly structure this finding. The issue is now here NVIDIA/cuopt#2042.