Describe the bug
On a large LP (7.2M rows × 2.27M cols, 14.3M nnz), enabling crossover after a
method=1 (PDLP) solve fails reproducibly with a numerical error late in the
crossover phase:
- Run 1: completes all 51,561 dual pushes, then aborts around push ~53.8k
of 101,781 primal pushes, after 2 h 32 min total —
Crossover status A numerical error was encountered., termination status 6,
nan primal/dual objectives.
- Run 2 (same structure, different variable bounds): completes all 44,975
dual pushes, then the same error during the primal pushes, after 2 h 50 min.
The PDLP phase itself converges fine on both runs (relative gap ~1e-4 / 1e-5
in ~9 min on the GPU). The crossover phase runs single-threaded on CPU at
~100 % of one core with the GPU idle, which makes the failure expensive to
discover: ~2.5 h of compute before the nan.
Steps/Code to reproduce bug
settings = SolverSettings()
settings.set_parameter("time_limit", 14400)
settings.set_parameter("method", 1) # PDLP
settings.set_parameter("crossover", 1)
solution = Solve(data_model, settings) # LP: 7.2M rows x 2.27M cols, 14.3M nnz
# -> "Crossover status A numerical error was encountered.", status 6, nan objectivesHappy to share full solver logs. (The instance itself is large; I can try to
provide it or a reproducer on request.)
Expected behavior
Either a successful crossover to a basic solution, or a structured/early
failure — ideally with a partial-progress diagnostic (e.g. remaining primal
infeasibilities) instead of nan objectives after ~2.5 h of pushes.
Environment details
- cuOpt version: 26.04.000 (git d9b7c96)
- Install method: Docker,
nvidia/cuopt:latest-cuda12.9-py3.13
- CUDA 12.9, driver-side GPU: NVIDIA GeForce RTX 3060 (12 GB)
- CPU: AMD Ryzen 5 8500G (6C/12T), 31 GB RAM
- OS: Ubuntu, kernel 6.8
Describe the bug
On a large LP (7.2M rows × 2.27M cols, 14.3M nnz), enabling
crossoverafter amethod=1(PDLP) solve fails reproducibly with a numerical error late in thecrossover phase:
of 101,781 primal pushes, after 2 h 32 min total —
Crossover status A numerical error was encountered., termination status 6,nanprimal/dual objectives.dual pushes, then the same error during the primal pushes, after 2 h 50 min.
The PDLP phase itself converges fine on both runs (relative gap ~1e-4 / 1e-5
in ~9 min on the GPU). The crossover phase runs single-threaded on CPU at
~100 % of one core with the GPU idle, which makes the failure expensive to
discover: ~2.5 h of compute before the
nan.Steps/Code to reproduce bug
Happy to share full solver logs. (The instance itself is large; I can try to
provide it or a reproducer on request.)
Expected behavior
Either a successful crossover to a basic solution, or a structured/early
failure — ideally with a partial-progress diagnostic (e.g. remaining primal
infeasibilities) instead of
nanobjectives after ~2.5 h of pushes.Environment details
nvidia/cuopt:latest-cuda12.9-py3.13