Add support for solver cuPDLPx (#516)

* Implement interface to cupdlpx solver and required IO

* Update tests

* Add to docs & add package dependency

* Add comment for release notes

* Correction: add support for equality constraints (same l & u bounds)

* Parameterize solution tolerance in tests to allow different standard for CPU and GPU solver precisions

* Improve cuPDLPx solver integration

- Add warning when explicit_coordinate_names is set (unsupported)
- Add warning when log_fn is provided (unsupported)
- Fix typo in comment (cuDPLPx -> cuPDLPx)
- Use pytest.skip() for clearer MIP test reporting
- Guard feasible_mip_solvers.remove() when cupdlpx unavailable

* Add cuPDLPx to package registry & update tests to use pattern for new solver features

* Add dynamic checking of package version when setting xpress solver features

* make xpress import save; use GPU/CPU_SOL_TOL

* update docs

* update gpu doc and remove cupdlx from solver deps

* fix: handle None constraint matrix in to_cupdlpx

* doc: add experimental warning to GPU acceleration docs

---------

Co-authored-by: Fabian Hofmann <fab.hof@gmx.de>

Author: mal84emma

CLAUDE.md

@@ -6,7 +6,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ### Running Tests
 ```bash
-# Run all tests
+# Run all tests (excluding GPU tests by default)
 pytest
 
 # Run tests with coverage
@@ -17,8 +17,16 @@ pytest test/test_model.py
 
 # Run a specific test function
 pytest test/test_model.py::test_model_creation
+
+# Run GPU tests (requires GPU hardware and cuPDLPx installation)
+pytest --run-gpu
+
+# Run only GPU tests
+pytest -m gpu --run-gpu
 ```
 
+**GPU Testing**: Tests that require GPU hardware (e.g., cuPDLPx solver) are automatically skipped by default since CI machines typically don't have GPUs. To run GPU tests locally, use the `--run-gpu` flag. The tests are automatically marked with `@pytest.mark.gpu` based on solver capabilities.
+
 ### Linting and Type Checking
 ```bash
 # Run linter (ruff)

README.md

@@ -149,6 +149,7 @@ Fri    0          4
 * [Cplex](https://www.ibm.com/de-de/analytics/cplex-optimizer)
 * [MOSEK](https://www.mosek.com/)
 * [COPT](https://www.shanshu.ai/copt)
+* [cuPDLPx](https://github.com/MIT-Lu-Lab/cuPDLPx)
 
 Note that these do have to be installed by the user separately.
 

doc/contributing.rst

@@ -10,6 +10,9 @@ contributing code.
 You are invited to submit pull requests / issues to our
 `Github repository <https://github.com/pypsa/linopy>`_.
 
+Development Setup
+=================
+
 For linting, formatting and checking your code contributions
 against our guidelines (e.g. we use `Black <https://github.com/psf/black>`_ as code style
 and use `pre-commit <https://pre-commit.com/index.html>`_:
@@ -19,6 +22,52 @@ and use `pre-commit <https://pre-commit.com/index.html>`_:
     * To automatically activate ``pre-commit`` on every ``git commit``: Run ``pre-commit install``
     * To manually run it: ``pre-commit run --all``
 
+Running Tests
+=============
+
+Testing is essential for maintaining code quality. We use pytest as our testing framework.
+
+Basic Testing
+-------------
+
+To run the test suite:
+
+.. code-block:: bash
+
+    # Install development dependencies
+    pip install -e .[dev,solvers]
+
+    # Run all tests
+    pytest
+
+    # Run tests with coverage
+    pytest --cov=./ --cov-report=xml linopy --doctest-modules test
+
+    # Run a specific test file
+    pytest test/test_model.py
+
+    # Run a specific test function
+    pytest test/test_model.py::test_model_creation
+
+GPU Testing
+-----------
+
+Tests for GPU-accelerated solvers (e.g., cuPDLPx) are automatically skipped by default since CI machines and most development environments don't have GPU hardware. This ensures tests pass in all environments.
+
+To run GPU tests locally (requires GPU hardware and CUDA):
+
+.. code-block:: bash
+
+    # Run all tests including GPU tests
+    pytest --run-gpu
+
+    # Run only GPU tests
+    pytest -m gpu --run-gpu
+
+GPU tests are automatically detected based on solver capabilities - no manual marking is required. When you add a new GPU solver to linopy, tests using that solver will automatically be marked as GPU tests.
+
+See the :doc:`gpu-acceleration` guide for more information about GPU solver setup and usage.
+
 Contributing examples
 =====================
 

doc/gpu-acceleration.rst

@@ -0,0 +1,117 @@
+========================
+GPU-Accelerated Solving
+========================
+
+.. warning::
+
+   This feature is **experimental** and not tested in CI due to the lack of GPU-enabled machines. Use with caution and please report any issues.
+
+Linopy supports GPU-accelerated optimization solvers that can significantly speed up solving large-scale linear programming problems by leveraging the parallel processing capabilities of modern GPUs.
+
+Supported GPU Solvers
+=====================
+
+cuPDLPx
+-------
+
+`cuPDLPx <https://github.com/MIT-Lu-Lab/cuPDLPx>`_ is an open-source, GPU-accelerated first-order solver developed by MIT. It implements a Primal-Dual hybrid gradient (PDHG) method optimized for GPUs.
+
+To install it, you have to have the `CUDA Toolkit <https://developer.nvidia.com/cuda/toolkit>`_ installed requiring NVIDIA GPUs on your computer. Then, install with
+
+.. code-block:: bash
+
+    # Install CUDA Toolkit first (if not already installed)
+    # Follow instructions at: https://developer.nvidia.com/cuda-downloads
+
+    # Install cuPDLPx
+    pip install cupdlpx>=0.1.2
+
+**Features:**
+
+- GPU-accelerated solving for large-scale linear programs
+- Open source (Apache 2.0 license)
+- Direct API integration with linopy
+- Designed for problems with millions of variables and constraints
+
+**Limitations:**
+
+- Currently supports only Linear Programming (LP)
+- Does not support Mixed-Integer Programming (MIP) or Quadratic Programming (QP)
+- Lower numerical precision compared to CPU solvers (typical tolerance: ~2.5e-4 vs 1e-5)
+- File I/O not currently supported through cuPDLPx API
+
+For a complete list of cuPDLPx parameters, see the `cuPDLPx documentation <https://github.com/MIT-Lu-Lab/cuPDLPx/tree/main/python#parameters>`_.
+
+Xpress with GPU Acceleration
+-----------------------------
+
+`FICO Xpress <https://www.fico.com/en/fico-xpress-trial-and-licensing-options>`_ version 9.8 and later includes GPU acceleration support for certain operations.
+
+**Features:**
+
+- Commercial solver with GPU support
+- Supports LP, MIP, and QP
+- Full-precision solving
+
+Prerequisites
+=============
+
+Hardware Requirements
+---------------------
+
+GPU solvers require:
+
+- NVIDIA GPU with CUDA support (compute capability 6.0 or higher recommended)
+- Sufficient GPU memory for your problem size (varies by problem)
+- PCIe 3.0 or higher for optimal data transfer
+
+Software Requirements
+---------------------
+
+1. **CUDA Toolkit**: Most GPU solvers require CUDA 11.0 or later
+2. **Compatible GPU drivers**: Match your CUDA version
+
+Verifying Installation
+======================
+
+To verify that the GPU solvers are properly installed and detected:
+
+.. code-block:: python
+
+    import linopy
+    from linopy.solver_capabilities import (
+        SolverFeature,
+        get_available_solvers_with_feature,
+    )
+
+    # Check available solvers
+    print("All available solvers:", linopy.available_solvers)
+
+    # Check GPU-accelerated solvers
+    gpu_solvers = get_available_solvers_with_feature(
+        SolverFeature.GPU_ACCELERATION, linopy.available_solvers
+    )
+    print("GPU solvers:", gpu_solvers)
+
+
+By default, GPU tests are skipped in the test suite to support CI environments without GPUs. To run GPU tests locally:
+
+.. code-block:: bash
+
+    # Run all tests including GPU tests
+    pytest --run-gpu
+
+    # Run only GPU tests
+    pytest -m gpu --run-gpu
+
+    # Run specific GPU test
+    pytest test/test_optimization.py -k cupdlpx --run-gpu
+
+
+References
+==========
+
+- `cuPDLPx Repository <https://github.com/MIT-Lu-Lab/cuPDLPx>`_
+- `cuPDLPx Python Documentation <https://github.com/MIT-Lu-Lab/cuPDLPx/tree/main/python>`_
+- `CUDA Installation Guide <https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html>`_
+- `NVIDIA GPU Computing Resources <https://developer.nvidia.com/gpu-computing>`_

doc/index.rst

@@ -37,7 +37,8 @@ flexible data-handling features:
 -  Use **lazy operations** for large linear programs with
    `dask <https://dask.org/>`__
 -  Choose from **different commercial and non-commercial solvers**
--  Fast **import and export** a linear model using xarray’s netcdf IO
+-  Fast **import and export** a linear model using xarray's netcdf IO
+-  Support for **GPU-accelerated solving** for large-scale problems
 -  Support of various solvers
    - `Cbc <https://projects.coin-or.org/Cbc>`__
    - `GLPK <https://www.gnu.org/software/glpk/>`__
@@ -48,6 +49,7 @@ flexible data-handling features:
    - `Cplex <https://www.ibm.com/de-de/analytics/cplex-optimizer>`__
    - `MOSEK <https://www.mosek.com/>`__
    - `COPT <https://www.shanshu.ai/copt>`__
+   - `cuPDLPx <https://github.com/MIT-Lu-Lab/cuPDLPx>`__ (GPU-accelerated)
 
 
 
@@ -116,6 +118,7 @@ This package is published under MIT license.
    infeasible-model
    solve-on-remote
    solve-on-oetc
+   gpu-acceleration
    migrating-from-pyomo
    gurobi-double-logging
 

doc/prerequisites.rst

@@ -30,11 +30,14 @@ Install a solver
 
 Linopy won't work without a solver. Currently, the following solvers are supported:
 
+CPU-based solvers
+~~~~~~~~~~~~~~~~~
+
 -  `Cbc <https://projects.coin-or.org/Cbc>`__ - open source, free, fast
 -  `GLPK <https://www.gnu.org/software/glpk/>`__ - open source, free, not very fast
 -  `HiGHS <https://www.maths.ed.ac.uk/hall/HiGHS/>`__ - open source, free, fast
 -  `Gurobi <https://www.gurobi.com/>`__  - closed source, commercial, very fast
--  `Xpress <https://www.fico.com/en/fico-xpress-trial-and-licensing-options>`__ - closed source, commercial, very fast
+-  `Xpress <https://www.fico.com/en/fico-xpress-trial-and-licensing-options>`__ - closed source, commercial, very fast (GPU acceleration available in v9.8+)
 -  `Cplex <https://www.ibm.com/de-de/analytics/cplex-optimizer>`__ - closed source, commercial, very fast
 -  `MOSEK <https://www.mosek.com/>`__
 -  `MindOpt <https://solver.damo.alibaba.com/doc/en/html/index.html>`__ -
@@ -54,6 +57,20 @@ We recommend to install the HiGHS solver if possible, which is free and open sou
     pip install highspy
 
 
+GPU-accelerated solvers
+~~~~~~~~~~~~~~~~~~~~~~~
+
+For large-scale optimization problems, GPU-accelerated solvers can provide significant performance improvements:
+
+-  `cuPDLPx <https://github.com/MIT-Lu-Lab/cuPDLPx>`__ - open source, GPU-accelerated first-order solver
+
+**Note:** GPU solvers require compatible NVIDIA GPU hardware and CUDA installation. See the :doc:`gpu-acceleration` guide for detailed setup instructions.
+
+.. code:: bash
+
+    pip install cupdlpx
+
+
 For most of the other solvers, please click on the links to get further installation information.
 
 

doc/release_notes.rst

@@ -10,6 +10,7 @@ Release Notes
 * Fix compatibility for xpress versions below 9.6 (regression)
 * Performance: Up to 50x faster ``repr()`` for variables/constraints via O(log n) label lookup and direct numpy indexing
 * Performance: Up to 46x faster ``ncons`` property by replacing ``.flat.labels.unique()`` with direct counting
+* Add support for GPU-accelerated solver [cuPDLPx](https://github.com/MIT-Lu-Lab/cuPDLPx)
 
 Version 0.5.8
 --------------

linopy/io.py

@@ -8,6 +8,7 @@
 import logging
 import shutil
 import time
+import warnings
 from collections.abc import Callable
 from io import BufferedWriter
 from pathlib import Path
@@ -28,6 +29,7 @@
 from linopy.objective import Objective
 
 if TYPE_CHECKING:
+    from cupdlpx import Model as cupdlpxModel
     from highspy.highs import Highs
 
     from linopy.model import Model
@@ -852,6 +854,72 @@ def to_highspy(m: Model, explicit_coordinate_names: bool = False) -> Highs:
     return h
 
 
+def to_cupdlpx(m: Model, explicit_coordinate_names: bool = False) -> cupdlpxModel:
+    """
+    Export the model to cupdlpx.
+
+    This function does not write the model to intermediate files but directly
+    passes it to cupdlpx.
+
+    cuPDLPx does not support named variables and constraints, so the
+    `explicit_coordinate_names` parameter is ignored.
+
+    Parameters
+    ----------
+    m : linopy.Model
+    explicit_coordinate_names : bool, optional
+        Ignored. cuPDLPx does not support named variables/constraints.
+
+    Returns
+    -------
+    model : cupdlpx.Model
+    """
+    import cupdlpx
+
+    if explicit_coordinate_names:
+        warnings.warn(
+            "cuPDLPx does not support named variables/constraints. "
+            "The explicit_coordinate_names parameter is ignored.",
+            UserWarning,
+            stacklevel=2,
+        )
+
+    # build model using canonical form matrices and vectors
+    # see https://github.com/MIT-Lu-Lab/cuPDLPx/tree/main/python#modeling
+    M = m.matrices
+    if M.A is None:
+        msg = "Model has no constraints, cannot export to cuPDLPx."
+        raise ValueError(msg)
+    A = M.A.tocsr()  # cuPDLPx only supports CSR sparse matrix format
+    # linopy stores constraints as Ax ?= b and keeps track of inequality
+    # sense in M.sense. Convert to separate lower and upper bound vectors.
+    l = np.where(
+        np.logical_or(np.equal(M.sense, ">"), np.equal(M.sense, "=")),
+        M.b,
+        -np.inf,
+    )
+    u = np.where(
+        np.logical_or(np.equal(M.sense, "<"), np.equal(M.sense, "=")),
+        M.b,
+        np.inf,
+    )
+
+    cu_model = cupdlpx.Model(
+        objective_vector=M.c,
+        constraint_matrix=A,
+        constraint_lower_bound=l,
+        constraint_upper_bound=u,
+        variable_lower_bound=M.lb,
+        variable_upper_bound=M.ub,
+    )
+
+    # change objective sense
+    if m.objective.sense == "max":
+        cu_model.ModelSense = cupdlpx.PDLP.MAXIMIZE
+
+    return cu_model
+
+
 def to_block_files(m: Model, fn: Path) -> None:
     """
     Write out the linopy model to a block structured output.

linopy/model.py

@@ -50,6 +50,7 @@
 )
 from linopy.io import (
     to_block_files,
+    to_cupdlpx,
     to_file,
     to_gurobipy,
     to_highspy,
@@ -1661,4 +1662,6 @@ def reset_solution(self) -> None:
 
     to_highspy = to_highspy
 
+    to_cupdlpx = to_cupdlpx
+
     to_block_files = to_block_files