We welcome all contributors to CVXPY. You don’t need to be an expert in convex optimization to help out. Here are simple ways to start contributing immediately:

If you’d like to add a new example to our library, or implement a new feature, please get in touch with us first by opening a GitHub issue to make sure that your priorities align with ours. We’ve identified some specific development opportunities in the next section.

CVXPY contributors must follow our code of conduct. Overall development is guided by our governance structure.

The rest of this page goes into more detail on how to contribute to CVXPY.


Here is a non-exhaustive list of opportunities to make prominent contributions to CVXPY. We’ve roughly categorized the opportunities as whether they amount to small-, medium-, or large-scope projects. New contributors are encouraged to focus on the projects with small or medium scope. Please contact a project maintainer if you’re interested in working on a project with large scope.

Small scope projects
Medium scope projects
Large scope projects

General principles

Development environment

Start by forking the CVXPY repository and installing CVXPY from source. You should configure git on your local machine before changing any code. Here’s one way CVXPY contributors might configure git:

  1. Tell git about the existence of the official CVXPY repo:

git remote add upstream
  1. Fetch a copy of the official master branch:

    git fetch upstream master
  2. Create a local branch which will track the official master branch:

    git branch --track official_master upstream/master

The only command you should use on the official_master branch is git pull. The purpose of this tracking branch is to allow you to easily sync with the main CVXPY repository. Such an ability can be a huge help in resolving any merge conflicts encountered in a pull request. For simple contributions, you might never use this branch.

  1. Switch back to your forked master branch:

    git checkout master
  2. Resume work as usual!

Contribution checklist

Contributions are made through pull requests. Before sending a pull request, make sure you do the following:

  • Add our license to new files

  • Check that your code adheres to our coding style.

  • Write unittests.

  • Run the unittests and check that they’re passing.

  • Run the benchmarks to make sure your change doesn’t introduce a regression

Once you’ve made your pull request, a member of the CVXPY development team will assign themselves to review it. You might have a few back-and-forths with your reviewer before it is accepted, which is completely normal. Your pull request will trigger continuous integration tests for many different Python versions and different platforms. If these tests start failing, please fix your code and send another commit, which will re-trigger the tests.


Please add the following license to new files:

Copyright, the CVXPY authors

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
See the License for the specific language governing permissions and
limitations under the License.

Code style

We use ruff to enforce our Python coding style. Before sending us a pull request, navigate to the project root and run

pip install ruff
ruff check cvxpy

to make sure that your changes abide by our style conventions. Please fix any errors that are reported before sending the pull request.

Optionally, the package pre-commit can be installed to check these conventions automatically before every commit.

pip install pre-commit
pre-commit install

Writing unit tests

Most code changes will require new unit tests. (Even bug fixes require unit tests, since the presence of bugs usually indicates insufficient tests.) CVXPY tests live in the directory cvxpy/tests, which contains many files, each of which contains many unit tests. When adding tests, try to find a file in which your tests should belong; if you’re testing a new feature, you might want to create a new test file.

We use the standard Python unittest framework for our tests. Tests are organized into classes, which inherit from BaseTest (see cvxpy/tests/ Every method beginning with test_ is a unit test.

Running unit tests

We use pytest to run our unit tests, which you can install with pip install pytest. To run all unit tests, cd into cvxpy/tests and run the following command:


To run tests in a specific file (e.g.,, use


To run a specific test method (e.g., TestDgp.test_product), use


Please make sure that your change doesn’t cause any of the unit tests to fail.

pytest suppresses stdout by default. To see stdout, pass the -s flag to pytest.


CVXPY has a few benchmarks in cvxpy/tests/, which test the time to canonicalize problems. Please run

pytest -s

with and without your change, to make sure no performance regressions are introduced. If you are making a code contribution, please include the output of the above command (with and without your change) in your pull request.

Solver interfaces

Third-party numerical optimization solvers are the lifeblood of CVXPY. We are very grateful to anyone who would be willing to volunteer their time to improve our existing solver interfaces, or create interfaces to new solvers. Improving an existing interface can usually be handled like fixing a bug. Creating a new interface requires much more work, and warrants coordination with CVXPY principal developers before writing any code.

This section of the contributing guide outlines considerations when adding new solver interfaces. For the time being, we only have documentation for conic solver interfaces. Additional documentation for QP solver interfaces is forthcoming.


This documentation is far from complete! It only tries to cover the absolutely essential parts of writing a solver interface. It also might not do that in a spectacular way – we welcome all feedback on this part of the documentation.


The developers try to keep this documentation up to date, however at any given time it might contain inaccurate information! It’s very important that you contact the CVXPY developers before writing a solver interface, if for no other reason than to prompt us to double-check the accuracy of this guide.

Conic solvers

Conic solvers require that the objective is a linear function of the optimization variable; constraints must be expressed using convex cones and affine functions of the optimization variable. The codepath for conic solvers begins with reductions/solvers/conic_solvers and in particular with the class ConicSolver in

Let’s say you’re writing a CVXPY interface for the “Awesome” conic solver, and that there’s an existing package AwesomePy for calling Awesome from python. In this case you need to create a file called in the same folder as Within you will define a class Awesome(ConicSolver). The Awesome(ConicSolver) class will manage all interaction between CVXPY and the existing AwesomePy python package. It will need to implement six functions:

  • import_solver,

  • name,

  • accepts,

  • apply,

  • solve_via_data, and

  • invert.

The first three functions are very easy (often trivial) to write. The remaining functions are called in order: apply stages data for solve_via_data, solve_via_data calls the Awesome solver by way of the existing third-party AwesomePy package, and invert transforms the output from AwesomePy into the format that CVXPY expects.

Key goals in this process are that the output of apply should be as close as possible to the Awesome’s standard form, and that solve_via_data should be kept short. The complexity of Awesome(ConicSolver).solve_via_data will depend on AwesomePy. If AwesomePy allows very low level input– passed by one or two matrices, and a handful of numeric vectors –then you’ll be in a situation like ECOS or GLPK. If the AwesomePy package requires that you build an object-oriented model, then you’re looking at something closer to the MOSEK, GUROBI, or NAG interfaces. Writing the invert function may require nontrivial effort to properly recover dual variables.

CVXPY’s conic form

CVXPY converts an optimization problem to an explicit form at the last possible moment. When CVXPY presents a problem in a concrete form, it’s over a single vectorized optimization variable, and a flattened representation of the feasible set. The abstraction for the standard form is

\[(P) \quad \min\{ c^T x + d \,:\, x \in \mathbb{R}^{n},\, A x + b \in K \}\]

where \(K\) is a product of elementary convex cones. The design of CVXPY allows for any cone supported by a target solver, but the current elementary convex cones are

  1. The zero cone \(y = 0 \in \mathbb{R}^m\).

  2. The nonnegative cone \(y \geq 0 \in \mathbb{R}^m\).

  3. The second order cone

    \[(u,v) \in K_{\mathrm{soc}}^n \doteq \{ (t,x) \,:\, t \geq \|x\|_2 \} \subset \mathbb{R} \times \mathbb{R}^n.\]
  4. One of several vectorized versions of the positive semidefinite cone.

  5. The exponential cone

\[(u,v,w) \in K_e \doteq \mathrm{cl}\{(x,y,z) | z \geq y \exp(x/y), y>0\}.\]
  1. The 3-dimensional power cone, parameterized by a number \(\alpha\in (0, 1)\):

    \[(u,v) \in K_{\mathrm{pow}}^{\alpha} \doteq \{ (x,y,z) \,:\, x^{\alpha}y^{1-\alpha} \geq |z|, (x,y) \geq 0 \}.\]

We address the vectorization options for the semidefinite cones later. For now it’s useful to say that the Awesome(ConicSolver) class will access an explicit representation for problem \((P)\) in in apply, with a code snippet like

# from cvxpy.constraints import Zero, NonNeg, SOC, PSD, ExpCone, PowCone3D
#  ...
if not problem.formatted:
    problem = self.format_constraints(problem, self.EXP_CONE_ORDER)
constr_map = problem.constr_map
cone_dims = problem.cone_dims
c, d, A, b = problem.apply_parameters()

The variable constr_map is is a dict of lists of CVXPY Constraint objects. The dict is keyed by the references to CVXPY’s Zero, NonNeg, SOC, PSD, ExpCone, and PowCone3D classes. You will need to interact with these constraint classes during dual variable recovery. For the other variables in that code snippet …

  • c, d define the objective function c @ x + d, and

  • A, b, cone_dims define the abstractions \(A\), \(b\), \(K\) in problem \((P)\).

The first step in writing a solver interface is to understand the exact meanings of A, b, cone_dims, so that you can correctly build a primal problem using the third-party AwesomePy interface to the Awesome solver. The cone_dims object is an instance of the ConeDims class, as defined in; A is a SciPy sparse matrix, and b is a numpy ndarray with b.ndim == 1. The rows of A and entries of b are given in a very specific order, as described below.

  • Equality constraints are found in the first rows of A and entries of b. Letting eq =, the constraint is

    A[:eq, :] @ x + b[:eq] == 0.
  • Inequality constraints occur immediately after the equations. If for example ineq = cone_dims.nonneg then the feasible set has the constraint

    A[eq:eq + ineq, :] @ x + b[eq:eq + ineq] >= 0.
  • Second order cone (SOC) constraints are handled after inequalities. Here, cone_dims.soc is a list of integers rather than a single integer. Supposing cone_dims.soc[0] == 10, the first second order cone constraint appearing in this optimization problem would involve 10 rows of A and 10 entries of b. The SOC vectorization we use is given by \(K_{\mathrm{soc}}^n\) as defined above.

  • PSD constraints follow SOC constraints. For most solver interfaces it is a good idea to make a deliberate decision about how to handle the vectorization, which amounts to implementing Awesome(ConicSolver).psd_format_mat. If you do nothing, then the vectorization will behave as in ConicSolver.psd_format_mat, which takes a PSD constraint of order \(n\) and maps it to \(n^2\) rows of \(A\) and entries of \(b\). You can also borrow from SCS.psd_format_mat which maps an order \(n\) PSD constraint to \(n(n+1)/2\) suitably scaled rows of \(A\) and entries of \(b\), or MOSEK.psd_format_mat which behaves identically to SCS except for the scaling.

  • The next block of 3 * cone_dims.exp rows in A, b correspond to consecutive three-dimensional exponential cones, as defined by \(K_e\) above.

  • The final block of 3 * len(cone_dims.p3d) rows in A, b correspond to three-dimensional power cones defined by \(K_{\mathrm{pow}}^{\alpha}\), where the i-th triple of rows has alpha = cone_dims.p3d[i].

If Awesome supports nonlinear constraints like SOC, ExpCone, PSD, or PowCone3D, then it’s possible that you will need to transform data A, b in order to write these constraints in the form expected by AwesomePy. The most common situations are when AwesomePy parametrizes the second-order cone as \(K = \{ (x,t) \,:\, \|x\|\leq t \} \subset \mathbb{R}^n \times \mathbb{R}\), or when it parametrizes \(K_e \subset \mathbb{R}^3\) as some permutation of what we defined earlier.

An alternative conic form

Some conic solvers do not natively support problem formats like (P) described in the previous section. Instead, the solver requires problem statements like

\[(Dir) \quad \min\{ f^T z \,:\, z \in K,\, G z = h \}.\]

Problem (Dir) uses so-called “direct” conic constraints \(z \in K\). If you are writing an interface for a solver which works this way, you should use the Dualize reduction on the standard CVXPY problem data given in (P). Using the Dualize reduction will avoid introduction unnecessary slack variables for continuous problems, but it is not applicable for problems with integer constraints. Therefore if your solver supports integer constraints, make sure to also use the Slacks reduction for that code path.

The MOSEK interface uses both of the reductions mentioned above.

Dual variables

Dual variable extraction should be handled in Awesome(ConicSolver).invert. To perform this step correctly, it’s necessary to consider how CVXPY forms a Lagrangian for the primal problem \((P)\). Let’s say that the affine map \(Ax + b\) in the feasible set \(Ax + b \in K \subset \mathbb{R}^m\) is broken up into six blocks of sizes \(m_1,\ldots,m_6\) where the blocks correspond (in order) to zero-cone, nonnegative cone, second-order cone, vectorized PSD cone, exponential cone, and 3D power cone constraints. Then CVXPY defines the dual to \((P)\) by forming a Lagrangian

\[\mathcal{L}(x,\mu_1,\ldots,\mu_6) = c^T x - \sum_{i=i}^6 \mu_i^T (A_i x + b_i)\]

in dual variables \(\mu_1 \in \mathbb{R}^{m_1}\), \(\mu_2 \in \mathbb{R}^{m_2}_+\), and \(\mu_i \in K_i^* \subset \mathbb{R}^{m_i}\) for \(i \in \{3,4,5,6\}\). Here, \(K_i^*\) denotes the dual cone to \(K_i\) under the standard inner product.

More remarks on dual variables (particularly SOC dual variables) can be found in this comment on a GitHub thread.

Most concrete implementations of the ConicSolver class use a common set of helper functions for dual variable recovery, found in reductions/solvers/

Registering a solver

Correctly implementing Awesome(ConicSolver) isn’t enough to call Awesome from CVXPY. You need to make edits in a handful of other places, namely

The existing content of those files should make it clear what’s needed to add Awesome to CVXPY.

Writing tests

Tests for Awesome(ConicSolver) should be placed in cvxpy/tests/ The overwhelming majority of tests in that file only take a single line, because we make consistent use of a general testing framework defined in Here are examples of helper functions we invoke in,

class StandardTestSDPs(object):

    def test_sdp_1min(solver, places=4, **kwargs):
        sth = sdp_1('min')
        sth.solve(solver, **kwargs)
        sth.verify_objective(places=2)  # only 2 digits recorded.
        sth.check_dual_domains(places)  # check dual variables are PSD.


class StandardTestSOCPs(object):

    def test_socp_0(solver, places=4, **kwargs):
        sth = socp_0()
        sth.solve(solver, **kwargs)


    def test_mi_socp_1(solver, places=4, **kwargs):
        sth = mi_socp_1()
        sth.solve(solver, **kwargs)
        # mixed integer problems don't have dual variables,
        #   so we only check the optimal objective and primal variables.

Notice the comments in the predefined functions. In test_sdp_1min, we override a user-supplied value for places with places=2 when checking the optimal objective function value. We also go through extra effort to check that the dual variables are PSD matrices. In test_mi_socp_1 we’re working with a mixed-integer problem, so there are no dual variables at all. You should use these predefined functions partly because they automatically check what’s most appropriate for the problem at hand.

Each of these predefined functions first constructs a SolverTestHelper object sth which contains appropriate test data. The .solve function for the SolverTestHelper class is a simple wrapper around prob.solve where prob is a CVXPY Problem. In particular, any keyword arguments passed to sth.solve will be passed to prob.solve. This allows you to call modified versions of a test with different solver parameters, for example

def test_mosek_lp_1(self):
    # default settings
    StandardTestLPs.test_lp_1(solver='MOSEK')  # 4 places
    # require a basic feasible solution
    StandardTestLPs.test_lp_1(solver='MOSEK', places=6, bfs=True)