Build against cuOpt 26.10 nightly: multi-GPU PDLP, new options, iteration/node reporting

[0x17]

• Build both workflows against a cuOpt 26.10 nightly from the RAPIDS nightly index, currently pinned to libcuopt-cuXX==26.10.0a223.post261002122804 as a stopgap
  • needs --prerelease=allow and --index-strategy unsafe-best-match so RAPIDS packages come from the nightly index and CUDA libraries from pypi.nvidia.com
  • a223 is the last nightly before cuOpt switched to modular component wheels (a224+). Against the modular a226 wheels the link builds, but every cuOptGetIntegerParameter/cuOptSetIntegerParameter call returns CUOPT_INVALID_ARGUMENT, so every solve ends with status 13
  • pins libcuopt-cuXX directly: there is no cuopt-cuXX wheel for a223, and pinning cuopt-cuXX==26.10.0a222 still resolves libcuopt to the modular a226
  • packaging for the modular layout (8edd916) is reverted in 4855ada and kept in history; unpin and reapply it once working modular 26.10 packages from NVIDIA are on PyPI
• Adapt link and bundles to the split cuOpt libraries inside the libcuopt wheel (both workflows and build-link.sh)
  • link -lcuopt_mathopt, bundle libcuopt/lib64/libcuopt_mathopt.so + libcuopt_client.so
  • copy cuOpt's cuDSS threading layer libcuopt/lib64/libcudss_mtlayer_cuopt.so; do not bundle libcudss_mtlayer_gomp.so
  • 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
  • bug fix: gmscuopt.c creates the problem with cuOptCreateRangedProblem instead of cuOptCreateProblem, since cuOpt's multi-GPU path only reads constraint lower/upper bounds and saw 0 constraints (validation error with presolve 0)
  • skip initial primal/dual solutions for multi-GPU PDLP, cuOpt rejects them
  • README section on usage and limitations
• Update optcuopt.def to cuOpt 26.10
  • 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)
  • new method 4 (primal simplex) and barrier_dual_initial_point 2 (SeDuMi-style), fix changed default of mip_hyper_heuristic_related_vars_time_limit
  • add 18 new options (e.g. concurrent_nnz_cutoff, primal_simplex_pricing, mip_rens, mip_mutation, barrier regularization, Curtis-Reid scaling); sequence_solve left out (Python re-solve cache only)
• 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)
• Verified
  • pinned nightly 26.10.0a223 (CUDA 13, single RTX A1000): regression tests 10/10, gamslib test 73/73 match CPLEX
  • PR CI builds and packages CUDA 12 and CUDA 13 on x86_64 and ARM64 with the pin (no solves in CI)
  • earlier against 26.10.0a222: multi-GPU code path (method 1, num_gpus -1, runs even on 1 GPU): 33/35 gamslib LPs match CPLEX within 1e-3; egypt hits the iteration limit (also with single-GPU PDLP), indus89 doesn't converge within 120 s (single-GPU PDLP: optimal in 25 s)
  • iterations/nodes and nodlim checked on trnsport and cube
  • 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)

[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.