DuaLip v5.0.0: GPU-Native LP Solver Rewritten in PyTorch

DuaLip v5.0.0: GPU-Native LP Solver Rewritten in PyTorch

DuaLip v5.0.0 is a complete rewrite of the solver, replacing the Spark/Scala runtime (v1.0.0–v4.0.5) with a GPU-native, pure-PyTorch implementation. This delivers major performance improvements and a simpler, composable API.

Highlights

• GPU-native solver — Significant wall-clock speedups via native CUDA execution through PyTorch. No custom C++/CUDA kernels required.
• Distributed multi-GPU support — Scales horizontally via torch.distributed (NCCL). Communication cost depends only on the dual dimension, enabling extreme-scale problems.
• Operator-centric API — Three decoupled primitives (ObjectiveFunction, ProjectionMap, Maximizer) make it easy to define new LP formulations by implementing only a new objective; projections and the solve loop are reused.
• Nesterov-accelerated gradient ascent — Includes Lipschitz-based step sizing, Jacobi preconditioning, and optional γ decay for faster convergence.
• Warm start support — Initialize from a previous dual solution for faster convergence on recurring workloads.
• MLflow integration — Track dual objective, step sizes, hyperparameters, and convergence diagnostics out of the box.

Requirements

• Python ≥ 3.10
• PyTorch ≥ 2.6.0

Breaking Changes

This is a full API redesign. The Spark/Scala API from v4.0.5 and earlier is no longer supported. See the README and examples for
migration guidance.

Contributors

Gregory Dexter, Aida Rahmattalabi, Sanjana Garg, Qingquan Song, Zhipeng Wang

v4.0.5

^{Changelog generated by Shipkit Changelog Gradle Plugin}

4.0.5

• 2023-11-16 - 1 commit(s) by Kenneth Tay
• Bugfix for PolytopeProjection.checkOptimality. (#47)

v4.0.4

^{Changelog generated by Shipkit Changelog Gradle Plugin}

4.0.4

• 2023-10-13 - 2 commit(s) by Kenneth Tay
• Minor update to wiki. (#45)

v4.0.2

^{Changelog generated by Shipkit Changelog Gradle Plugin}

4.0.2

• 2023-07-25 - 1 commit(s) by Kenneth Tay
• Add spark session as an argument to a new method to run the solver. (#44)

v4.0.1

^{Changelog generated by Shipkit Changelog Gradle Plugin}

4.0.1

• 2023-07-25 - 1 commit(s) by Kenneth Tay
• Fix bug in twoStepGradientAggregator method. (#43)

v4.0.0

^{Changelog generated by Shipkit Changelog Gradle Plugin}

4.0.0

• 2023-07-25 - 4 commit(s) by Ayan Acharya, Borja Ocejo, Miao Cheng
• Add case classes to enforce matching problem budget data assumptions, Release DuaLip v4.0 (#42)
• order of local and global budgets got reversed in the initialization of constrained matching solver, fixing that issue (#41)

v3.0.2

^{Changelog generated by Shipkit Changelog Gradle Plugin}

3.0.2

• 2023-06-12 - 1 commit(s) by Ayan Acharya
• No notable improvements. No pull requests (issues) were referenced from commits.

v3.0.1

^{Changelog generated by Shipkit Changelog Gradle Plugin}

3.0.1

• 2023-06-08 - 2 commit(s) by Ayan Acharya, Kenneth Tay
• Rename preprocess/MatchingDataGenerator.scala for parity between versions. (#40)

v3.0.0

^{Changelog generated by Shipkit Changelog Gradle Plugin}

3.0.0

• 2023-05-19 - 4 commit(s) by Ayan Acharya, Kenneth Tay
• establish parity between external and internal solver (#39)
• Update Github README to match wiki. (#35)

v2.8.1

^{Changelog generated by Shipkit Changelog Gradle Plugin}

2.8.1

• 2023-04-19 - 1 commit(s) by Borja Ocejo
• Updated contributions.md and .gitignore (#34)