torchPDLP 0.1.2

Uses a PyTorch-implemented restarted PDHG algorithm with enhancements to solve primal-dual linear programs.

torchPDLP

torchPDLP is a PyTorch-based solver for linear programs, using a restarted PDHG algorithm with enhancements for stability and performance.

Features

• Efficient primal-dual LP solving with PyTorch acceleration
• Directory and single-file solving modes
• Outputs results and solver status in accessible formats
• Enhanced algorithmic stability and speed

Installation

Install from PyPI:

pip install torchPDLP

Or install directly from GitHub:

pip install git+https://github.com/SimplySnap/torchPDLP.git@pypi-package#subdirectory=torchPDLP

Usage

1. Solving a Directory of Problems from Command Line

You can use the provided command-line script to solve all MPS files in a directory and output a summary CSV file:

torchPDLP \
  --device gpu \
  --instance_path /path/to/mps/files \
  --tolerance 1e-4 \
  --output_path /path/to/save/results \
  --precondition \
  --fishnet \
  --primal_weight_update \
  --adaptive_stepsize \
  --max_kkt 100000

Argument Reference:

Argument	Description
--device	'cpu', 'gpu', or 'auto'. Uses GPU if available as default.
--instance_path	Path to folder with .mps files.
--tolerance	Convergence tolerance (default: 1e-4).
--output_path	Folder to save outputs and Excel results.
--precondition	Enable Ruiz preconditioning (optional).
--primal_weight_update	Enable primal weight updates (optional).
--adaptive_stepsize	Enable adaptive step sizes (optional).
--fishnet	Enable fishnet casting (optional).
--verbose	Enable verbose logging (optional).
--support_sparse	Use sparse matrices if supported (optional).
--max_kkt	Maximum number of KKT passes (default: None).

2. Solving a Single Problem in Python

To solve a single MPS problem file and retrieve the solution, objective value, and solver status:

import torchPDLP

result = torchPDLP.solve("path/to/file.mps")

torchPDLP.solve has all the same optional arguments as the command line function. Result is a dictionary with keys:

Key	Description
optimal_point	The optimal solution found by the solver (PyTorch tensor).
objective_value	The value of the objective function at the optimal point (float).
status	Solver status, either "solved" or "unsolved" (string).
time	Total time taken to solve the problem (in seconds, float).
iterations	Number of main algorithm iterations performed (integer).
restarts	Number of times the PDHG algorithm was restarted (integer).
kkt_passes	Number of KKT passes performed during solving (integer).

Authors

• Xiyan Hu, Colgate University
• Titus Parker, Stanford University
• Connor Phillips, James Madison University
• Yifa Yu, University of California, Davis

License

MIT License