Incorporate determinism into GPU heuristics + parallel determinsitic B&B+GPU exploration by aliceb-nv · Pull Request #986 · NVIDIA/cuopt

aliceb-nv · 2026-03-23T11:13:09Z

This PR implements the integration of GPU heuristics into the deterministic codepath of the solver. This enables running full end-to-end solves with presolve, parallel B&B, and GPU heuristics, with full run-to-run determinism and heterogeneous (CPU+GPU) execution.

For this purpose, the GPU heuristics were modified to eliminate sources of nondeterminism such as non-ordered floating point reductions, atomic operations, scheduling-dependent operations... Extensive testing has been added to the CI runs to ensure individual components (such as probing, FJ, recombiners, FP...) are bitwise deterministic, and their overall orchestration maintains this invariant.
Furthermore, to reduce changes to the codebase and splits in the codepath, wall-clock timers on the heuristics side were replaced with unified termination checking objects, that check for wall time (opportunistic mode), or work total (deterministic mode).
These timers are hierarchical, and thus allow for checking against multiple terminations sources, for example a local algo's work budget, the global wall clock time limit, or in future PRs user-controlled termination signals like Ctrl-C or callback handlers.

B&B acts as the authority for incumbent publishing to the user. At every horizon sync, the queue of GPU heuristic solutions is checked, and if solutions whose timestamp is <= the current B&B work total, they are drained, replayed, and incorporated into the tree exploration.

A new get_solution_callback_ext callback has been introduced to prevent ABI breaks and pass more metadata with each published solution. For now, only the solution origin and its work unit timestamp are exposed, but additional fields may be later added without breaking the ABI.
The benchmark runner now relies on such callbacks to record the incumbents and their timestamps - it is no longer necessary to parse the logs to compute the primal integral.

A few other bugfixes have been included to the existing B&B deterministic exploration regarding node queues. Other fixes include uninitialized memory accesses, and improvements to allow running compute-sanitizer initcheck without false positives. Determinism logs have also been added thorought the code, disabled by default.

Some features remain unsupported in deterministic mode, such as Papilo's Probing presolver, concurrent root solve, RINS, or the sub-MIP recombiner. These will be incorporated in later PRs.
Furthermore, the PR has been modified to take the ML estimators out for the sake of reducing code review workload. They will be incorporated again in later PRs, along with improved tuning.
No intentional API or ABI breaking changes introduced.

Opportunistic mode performance remains unchanged:
Primal gap 12.5%, 226 feasible, primal integral 19.1%.

Deterministic mode benchmarks pending.

Closes #144
Closes #882

github-actions · 2026-04-17T09:37:36Z

🔔 Hi @anandhkb, this pull request has had no activity for 7 days. Please update or let us know if it can be closed. Thank you!