DWave.jl

D-Wave Quantum Annealing Interface for JuMP

Installation

julia> import Pkg

julia> Pkg.add("DWave")

Migration from DWaveNeal.jl

DWaveNeal.jl only covered the simulated annealing wrapper that now lives at DWave.Neal.Optimizer inside DWave.jl. The rest of DWave.jl provides additional samplers and is not part of the old DWaveNeal.jl API surface.

For code that used DWaveNeal.Optimizer, install DWave.jl and switch to DWave.Neal.Optimizer:

julia> import Pkg

julia> Pkg.add("DWave")

julia> using DWave

If you still have an environment that depends on DWaveNeal, Pkg.add("DWaveNeal") continues to work, but DWaveNeal.Optimizer now aliases DWave.Neal.Optimizer and emits a deprecation warning on load.

Basic Usage

using JuMP
using QUBO
using DWave

model = Model(DWave.Neal.Optimizer)

h = [-1, -1, -1]
J = [0 2 2; 0 0 2; 0 0 0]

@variable(model, s[1:3], Spin)

@objective(model, Min, h's + s'J * s)

optimize!(model)

for i = 1:result_count(model)
    si = value.(s; result=i)
    yi = objective_value(model; result=i)

    println("H($si) = $yi")
end

Sampler Overview

DWave.jl provides two families of optimizers:

• DWave.Optimizer connects to D-Wave cloud samplers and requires access to a Leap account.
• DWave.Neal.Optimizer, DWave.Greedy.Optimizer, DWave.Random.Optimizer, and DWave.Tabu.Optimizer wrap the classical samplers shipped in dwave-samplers and run locally through the PythonCall/CondaPkg environment managed by the package.

Switching between them only requires changing the optimizer passed to Model(...).

Classical Samplers

The classical samplers expose the same QUBO/Ising modeling interface as the QPU wrapper, but they do not require DWAVE_API_TOKEN.

• DWave.Neal.Optimizer
• DWave.Greedy.Optimizer
• DWave.Random.Optimizer
• DWave.Tabu.Optimizer

Example:

using JuMP
using QUBO
using DWave

model = Model(DWave.Tabu.Optimizer)

set_attribute(model, "num_reads", 32)
set_attribute(model, "timeout", 100)
set_attribute(model, "initial_states", [
     1  1 -1 -1
    -1 -1  1  1
])

h = [-1, -1, 0, 0]
J = [0 0 1 0; 0 0 0 1; 0 0 0 0; 0 0 0 0]

@variable(model, s[1:4], Spin)
@objective(model, Min, h' * s + s' * J * s)

optimize!(model)

Sampler-specific options are forwarded as raw optimizer attributes via set_attribute(model, name, value). For initial_states, the Greedy and Tabu wrappers accept either a single Julia vector or a matrix whose rows are initial states.

Optimizer	Description	Raw optimizer attributes
DWave.Neal.Optimizer	Simulated annealing baseline.	num_reads, num_sweeps, num_sweeps_per_beta, beta_range, beta_schedule, beta_schedule_type, seed, initial_states_generator, interrupt_function
DWave.Greedy.Optimizer	Steepest-descent local search.	num_reads, initial_states, initial_states_generator, seed, large_sparse_opt
DWave.Random.Optimizer	Random-state sampling.	num_reads, time_limit, max_num_samples, seed
DWave.Tabu.Optimizer	Tabu-search local search.	initial_states, initial_states_generator, num_reads, seed, tenure, timeout, num_restarts, energy_threshold, coefficient_z_first, coefficient_z_restart, lower_bound_z

The upstream planar and tree samplers are not wrapped yet. PlanarGraphSolver only applies to planar Ising models without linear biases, and the tree decomposition samplers expose exact-solver and marginal APIs that do not fit the current QUBODrivers interface cleanly.

API Token

To use D-Wave's QPU it is necessary to obtain an API Token from Leap.

Topology and Embedding Plots

DWave.draw_topology and DWave.draw_embedding wrap D-Wave NetworkX's Pegasus and Zephyr plotting helpers and return a Matplotlib figure by default. Both helpers accept DWave.WorkingGraph values built from sampler metadata, so plots use the full calibrated working graph rather than only the embedded qubits.

using DWave
using QUBOTools

metadata = QUBOTools.metadata(sampleset)

DWave.draw_topology(DWave.WorkingGraph(metadata))
DWave.draw_embedding(metadata)

Pass ax = existing_axis to draw into an existing Matplotlib axis; all other keyword arguments are forwarded to the corresponding D-Wave NetworkX draw function.

Development Checks

To verify that DWave can share a CondaPkg environment with the other Python-backed JuliaQUBO benchmark drivers, run:

julia --startup-file=no test/shared_condapkg_resolve.jl

Disclaimer: The D-Wave wrapper for Julia is not officially supported by D-Wave Systems. If you are a commercial customer interested in official support for Julia from D-Wave, let them know!

Note: If you are using DWave.jl in your project, we recommend you to include the .CondaPkg entry in your .gitignore file. The PythonCall module will place a lot of files in this folder when building its Python environment.