# 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:
```bash
pip install torchPDLP
```
Or install directly from GitHub:
```bash
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:
```bash
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:
```python
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
Raw data
{
"_id": null,
"home_page": null,
"name": "torchPDLP",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.6",
"maintainer_email": null,
"keywords": "optimization, primal-dual, linear programming, PDHG, torch",
"author": "Xiyan Hu, Titus Parker, Connor Phillips, Yifa Yu",
"author_email": null,
"download_url": "https://files.pythonhosted.org/packages/f6/61/27a86e48bc7e8424fbb8b60e16d8dd9745e2ebd36146d6fbfd3702db4b23/torchpdlp-0.1.2.tar.gz",
"platform": null,
"description": "# torchPDLP\n\ntorchPDLP is a PyTorch-based solver for linear programs, using a restarted PDHG algorithm with enhancements for stability and performance.\n\n## Features\n\n- Efficient primal-dual LP solving with PyTorch acceleration\n- Directory and single-file solving modes\n- Outputs results and solver status in accessible formats\n- Enhanced algorithmic stability and speed\n\n## Installation\n\nInstall from PyPI:\n```bash\npip install torchPDLP\n```\n\nOr install directly from GitHub:\n```bash\npip install git+https://github.com/SimplySnap/torchPDLP.git@pypi-package#subdirectory=torchPDLP\n```\n\n## Usage\n\n### 1. Solving a Directory of Problems from Command Line\n\nYou can use the provided command-line script to solve all MPS files in a directory and output a summary CSV file:\n\n```bash\ntorchPDLP \\\n --device gpu \\\n --instance_path /path/to/mps/files \\\n --tolerance 1e-4 \\\n --output_path /path/to/save/results \\\n --precondition \\\n --fishnet \\\n --primal_weight_update \\\n --adaptive_stepsize \\\n --max_kkt 100000\n```\n Argument Reference:\n\n| Argument | Description |\n| ------------------------ | ---------------------------------------------------------------------------- |\n| `--device` | `'cpu'`, `'gpu'`, or `'auto'`. Uses GPU if available as default. |\n| `--instance_path` | Path to folder with `.mps` files. |\n| `--tolerance` | Convergence tolerance (default: `1e-4`). |\n| `--output_path` | Folder to save outputs and Excel results. |\n| `--precondition` | Enable Ruiz preconditioning (optional). |\n| `--primal_weight_update` | Enable primal weight updates (optional). |\n| `--adaptive_stepsize` | Enable adaptive step sizes (optional). |\n| `--fishnet` | Enable fishnet casting (optional). |\n| `--verbose` | Enable verbose logging (optional). |\n| `--support_sparse` | Use sparse matrices if supported (optional). |\n| `--max_kkt` | Maximum number of KKT passes (default: `None`). |\n\n### 2. Solving a Single Problem in Python\n\nTo solve a single MPS problem file and retrieve the solution, objective value, and solver status:\n\n```python\nimport torchPDLP\n\nresult = torchPDLP.solve(\"path/to/file.mps\")\n```\ntorchPDLP.solve has all the same optional arguments as the command line function.\nResult is a dictionary with keys:\n\n| Key | Description |\n| ----------------- | --------------------------------------------------------------------------------------------- |\n| `optimal_point` | The optimal solution found by the solver (PyTorch tensor). |\n| `objective_value` | The value of the objective function at the optimal point (float). |\n| `status` | Solver status, either `\"solved\"` or `\"unsolved\"` (string). |\n| `time` | Total time taken to solve the problem (in seconds, float). |\n| `iterations` | Number of main algorithm iterations performed (integer). |\n| `restarts` | Number of times the PDHG algorithm was restarted (integer). |\n| `kkt_passes` | Number of KKT passes performed during solving (integer). |\n\n\n## Authors\n\n- Xiyan Hu, Colgate University\n- Titus Parker, Stanford University\n- Connor Phillips, James Madison University\n- Yifa Yu, University of California, Davis\n\n## License\n\nMIT License\n",
"bugtrack_url": null,
"license": "MIT",
"summary": "Uses a PyTorch-implemented restarted PDHG algorithm with enhancements to solve primal-dual linear programs.",
"version": "0.1.2",
"project_urls": {
"Homepage": "https://github.com/SimplySnap/torchPDLP"
},
"split_keywords": [
"optimization",
" primal-dual",
" linear programming",
" pdhg",
" torch"
],
"urls": [
{
"comment_text": null,
"digests": {
"blake2b_256": "b0bbae8f7dfdefdbe1b90c70993e18b2f8bb96f161d04c3cc3734e67aa28f6c5",
"md5": "1c88bb5c378d4c0482fc9aff74af213b",
"sha256": "b4092040dcde1e2fe3bfe9463751cb3d2e6ec00c4aba36e4632b1f439fa132d2"
},
"downloads": -1,
"filename": "torchpdlp-0.1.2-py3-none-any.whl",
"has_sig": false,
"md5_digest": "1c88bb5c378d4c0482fc9aff74af213b",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.6",
"size": 33508,
"upload_time": "2025-08-20T02:39:07",
"upload_time_iso_8601": "2025-08-20T02:39:07.448699Z",
"url": "https://files.pythonhosted.org/packages/b0/bb/ae8f7dfdefdbe1b90c70993e18b2f8bb96f161d04c3cc3734e67aa28f6c5/torchpdlp-0.1.2-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": null,
"digests": {
"blake2b_256": "f66127a86e48bc7e8424fbb8b60e16d8dd9745e2ebd36146d6fbfd3702db4b23",
"md5": "81d0b24790045fa97621d327cc80a30c",
"sha256": "388373c2072faf06f4d4d9b442a3a8ef3aa304e5989f0908ff483b381f9560fb"
},
"downloads": -1,
"filename": "torchpdlp-0.1.2.tar.gz",
"has_sig": false,
"md5_digest": "81d0b24790045fa97621d327cc80a30c",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.6",
"size": 30702,
"upload_time": "2025-08-20T02:39:08",
"upload_time_iso_8601": "2025-08-20T02:39:08.514014Z",
"url": "https://files.pythonhosted.org/packages/f6/61/27a86e48bc7e8424fbb8b60e16d8dd9745e2ebd36146d6fbfd3702db4b23/torchpdlp-0.1.2.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2025-08-20 02:39:08",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "SimplySnap",
"github_project": "torchPDLP",
"travis_ci": false,
"coveralls": false,
"github_actions": false,
"lcname": "torchpdlp"
}
JSON
