Python

Python interface supports Python 3.8 or newer.

Pip

Install osqp with the default algebra backend using pip:

pip install osqp

The builtin algebra backend is always available for use. Alternative osqp algebra backends - mkl or cu12 as of the time of this writing, can also be installed:

To install osqp with the mkl backend:

pip install osqp[mkl]

To install osqp with the cu12 (Cuda 12.x) backend:

pip install osqp[cu12]

To install osqp with the mkl and cu12 backends:

pip install osqp[mkl,cu12]

Note

These commands install osqp with the mkl or cu12 “extras”, which provide the osqp-mkl or osqp-cuda packages respectively. These extension modules are directly importable using import osqp_mkl or import osqp_cuda, though you will never directly need to do this.

Using Alternative Algebra Backends

Checking which backends are installed

Use the algebras_available function to check what algebra backends you have available locally:

$ python -c "from osqp import algebras_available; print(algebras_available())"
['cuda', 'mkl', 'builtin']

Note

Installing alternative algebras like mkl or cu12 does not automatically install the runtime MKL or Cuda libraries for your system. If you have installed osqp with the mkl or cu12 extras but do not see the expected algebra(s) available, you may need to install these libraries separately, and ensure that they are available to the Python interpreter at runtime.

For example, to use mkl, you might need to run:

$ pip install mkl-devel && LD_LIBRARY_PATH=$LD_LIBRARY_PATH:.venv/lib

$ conda install conda-forge::mkl

To use cuda, you might need to install the Cuda toolkit for your system, or (more easily) for your python environment:

pip install nvidia-cublas-cu12 nvidia-cusparse-cu12 nvidia-cuda-runtime-cu12
LD_LIBRARY_PATH=$LD_LIBRARY_PATH:.venv/lib/nvidia/cusparse/lib

conda install conda-forge::cudatoolkit

The exact path to the libraries will vary depending on your system. If unsure, you can try to directly import the extension modules to see what additional runtime libraries are needed:

$ python -c "import osqp_mkl"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
ImportError: libmkl_rt.so.2: cannot open shared object file: No such file or directory

$ python -c "import osqp_cuda"

If you’re using osqp in a cluster environment, it might be sufficient to just activate a module that provides the necessary libraries. For example, on our clusters, we do a module load intel-mkl/2024.0 or module load cudatoolkit/12.6 to get the necessary libraries.

Switching algebra backends

Use the default_algebra function to check what algebra backend is currently set as the default:

$ python -c "from osqp import default_algebra; print(default_algebra())"
cuda

By default, osqp uses the best algebra that is available (cuda being preferred over mkl, which is preferred over builtin). The default algebra can be overridden by setting the OSQP_ALGEBRA_BACKEND environment variable to one of builtin, mkl, or cuda.

$ OSQP_ALGEBRA_BACKEND=builtin python -c "from osqp import default_algebra; print(default_algebra())"
builtin

You can set this environment variable in your shell, or in your python script, before importing osqp. Alternatively, you can specify the algebra argument to osqp.OSQP to use a specific algebra backend for a particular problem.

$ python -c "from osqp import OSQP; problem = OSQP(algebra='mkl'); print(problem.algebra)"
mkl

$ python -c "from osqp import OSQP; problem = OSQP(algebra='builtin'); print(problem.algebra)"
builtin