README.md

GPU implementation of Clarabel solver for Julia

Features • Installation • License • Documentation

CuClarabel.jl is the GPU implementation of the Clarabel solver, which can solve conic problems of the following form:

$$ \begin{array}{r} \text{minimize} & \frac{1}{2}x^T P x + q^T x\\[2ex] \text{subject to} & Ax + s = b \\[1ex] & s \in \mathcal{K} \end{array} $$

with decision variables $x \in \mathbb{R}^n$, $s \in \mathbb{R}^m$ and data matrices $P=P^\top \succeq 0$, $q \in \mathbb{R}^n$, $A \in \mathbb{R}^{m \times n}$, and $b \in \mathbb{R}^m$. The set $\mathcal{K}$ is a composition of convex cones; we support zero cones (linear equality constraints), nonnegative cones (linear inequality constraints), second-order cones, exponential cone and power cones. Our package relies on the external package CUDSS.jl for the linear system solver CUDSS. We also support linear system solves in lower (mixed) precision.

Installation

• CuClarabel.jl can be added via the Julia package manager (type ]): pkg> dev https://github.com/cvxgrp/CuClarabel.git, (which will overwrite current use of Clarabel solver).

Tutorial

Modeling a conic optimization problem is the same as in the original Clarabel solver, except with the additional parameter direct_solve_method. This can be set to :cudss or :cudssmixed. Here is a portfolio optimization problem modelled via JuMP:

using LinearAlgebra, SparseArrays, Random, JuMP
using Clarabel

## generate the data
rng = Random.MersenneTwister(1)
k = 5; # number of factors
n = k * 10; # number of assets
D = spdiagm(0 => rand(rng, n) .* sqrt(k))
F = sprandn(rng, n, k, 0.5); # factor loading matrix
μ = (3 .+ 9. * rand(rng, n)) / 100. # expected returns between 3% - 12%
γ = 1.0; # risk aversion parameter
d = 1 # we are starting from all cash
x0 = zeros(n);

a = 1e-3
b = 1e-1
γ = 1.0;

model = JuMP.Model(Clarabel.Optimizer)
set_optimizer_attribute(model, "direct_solve_method", :cudss)

@variable(model, x[1:n])
@variable(model, y[1:k])   
@variable(model, s[1:n])
@variable(model, t[1:n])
@objective(model, Min, x' * D * x + y' * y - 1/γ * μ' * x);
@constraint(model, y .== F' * x);
@constraint(model, x .>= 0);

# transaction costs
@constraint(model, sum(x) + a * sum(s) == d + sum(x0) );
@constraint(model, [i = 1:n], x0[i] - x[i] == t[i]) 
@constraint(model, [i = 1:n], [s[i], t[i]] in MOI.SecondOrderCone(2));
JuMP.optimize!(model)

Citing

@misc{CuClarabel,
      title={CuClarabel: GPU Acceleration for a Conic Optimization Solver}, 
      author={Yuwen Chen and Danny Tse and Parth Nobel and Paul Goulart and Stephen Boyd},
      year={2024},
      eprint={2412.19027},
      archivePrefix={arXiv},
      primaryClass={math.OC},
      url={https://arxiv.org/abs/2412.19027}, 
}

License 🔍

This project is licensed under the Apache License 2.0 - see the LICENSE.md file for details.