Installation

Quick Install

Most Users

pip install cvxpylayers[torch]

Includes PyTorch and all dependencies.

Minimal Install

pip install cvxpylayers

Core only — add framework extras as needed.

---

Choose Your Framework

PyTorch

pip install cvxpylayers[torch]

Requirements: PyTorch >= 2.0

The most popular choice. Full torch.nn.Module integration with autograd support.

JAX

pip install cvxpylayers[jax]

Requirements: JAX >= 0.4.0

Functional style with jax.grad, jax.vmap, and jax.jit (Moreau solver).

MLX

pip install cvxpylayers[mlx]

Requirements: MLX >= 0.27.1, Apple Silicon Mac

Optimized for M1/M2/M3 chips with unified memory.

All Frameworks

pip install cvxpylayers[all]

Install everything — useful for development or testing.

---

Verify Installation

PyTorch

import cvxpy as cp
import torch
from cvxpylayers.torch import CvxpyLayer

x = cp.Variable(2)
A = cp.Parameter((2, 2))
b = cp.Parameter(2)
problem = cp.Problem(cp.Minimize(cp.sum_squares(A @ x - b)))

layer = CvxpyLayer(problem, parameters=[A, b], variables=[x])
A_t = torch.eye(2, requires_grad=True)
b_t = torch.ones(2, requires_grad=True)
(sol,) = layer(A_t, b_t)

print(f"Solution: {sol}")  # tensor([1., 1.])
print("Installation successful!")

JAX

import cvxpy as cp
import jax.numpy as jnp
from cvxpylayers.jax import CvxpyLayer

x = cp.Variable(2)
A = cp.Parameter((2, 2))
b = cp.Parameter(2)
problem = cp.Problem(cp.Minimize(cp.sum_squares(A @ x - b)))

layer = CvxpyLayer(problem, parameters=[A, b], variables=[x])
A_jax = jnp.eye(2)
b_jax = jnp.ones(2)
(sol,) = layer(A_jax, b_jax)

print(f"Solution: {sol}")  # [1., 1.]
print("Installation successful!")

MLX

import cvxpy as cp
import mlx.core as mx
from cvxpylayers.mlx import CvxpyLayer

x = cp.Variable(2)
A = cp.Parameter((2, 2))
b = cp.Parameter(2)
problem = cp.Problem(cp.Minimize(cp.sum_squares(A @ x - b)))

layer = CvxpyLayer(problem, parameters=[A, b], variables=[x])
A_mx = mx.eye(2)
b_mx = mx.ones(2)
(sol,) = layer(A_mx, b_mx)

print(f"Solution: {sol}")  # [1., 1.]
print("Installation successful!")

---

GPU Acceleration

CuClarabel — NVIDIA GPU Support

For large-scale problems, CuClarabel keeps everything on GPU with no CPU transfers.

Step 1: Julia

Install from julialang.org

Step 2: Python

pip install juliacall cupy diffqcp

Step 3: CuClarabel

In Python run:

from juliacall import Main as jl 
jl.seval('using Pkg; Pkg.add(url="https://github.com/oxfordcontrol/Clarabel.jl", rev="CuClarabel")')

Usage:

import cvxpy as cp
from cvxpylayers.torch import CvxpyLayer

layer = CvxpyLayer(
    problem,
    parameters=[A, b],
    variables=[x],
    solver=cp.CUCLARABEL
).to("cuda")

# Parameters must be on GPU
A_gpu = A_t.cuda()
b_gpu = b_t.cuda()
(solution,) = layer(A_gpu, b_gpu)

---

Dependencies

Core Dependencies

Package	Version	Purpose
Python	>= 3.11	Runtime
NumPy	>= 1.22.4	Array operations
CVXPY	>= 1.7.4	Problem specification
diffcp	>= 1.1.0	Differentiable cone programming

Framework Dependencies

Framework	Version
PyTorch	>= 2.0
JAX	>= 0.4.0
MLX	>= 0.27.1

---

Development Setup

For contributing:

git clone https://github.com/cvxpy/cvxpylayers.git
cd cvxpylayers
pip install -e ".[all,dev]"

Run tests:

pytest tests/