Ascent Science Python SDK

Ascent Science SDK 0.7.0 is installed as one complete distribution. Algorithms are organized by type and selected through Python subpackages:

python -m pip install --upgrade ascent-science==0.7.0

import ascent

aco = ascent.combinatorial.ACOOptimizer(ants=30, iterations=100)
pso = ascent.combinatorial.PSOOptimizer(particles=30, iterations=100)
exact = ascent.qubo.ExactQUBOOptimizer()
simulator = ascent.qubo.CIMSimulator()
hardware = ascent.cim.CIMOptimizer(task_name="production-qubo")

Algorithm namespaces follow the platform algorithm repository exactly:

Only implemented and packaged algorithms are exposed as callable classes. Prototype and planned entries shown in the web directory are not represented as working SDK classes. ascent.qubo, ascent.cim, and ascent.core are cross-category functional modules rather than repository algorithm categories.

The SDK also separates problem models, solver capabilities, and two execution modes:

• ascent.core provides stable Problem, Solver, SolveResult, registry, and QUBO/Ising transformation contracts.

• QuboProblem, IsingProblem, TSPProblem, and ContinuousProblem decouple a problem from the selected solver.

• Category optimizer classes execute local algorithms in the caller's Python process.

• ascent.cim.CIMOptimizer submits QUBO/Ising matrices to an explicitly configured real CIM gateway, polls task state, downloads the result, and caches successful results locally.

The hardware client never falls back to a simulator. Without ASCENT_CIM_ENDPOINT, construction raises ConfigurationError.

Local execution

import ascent

optimizer = ascent.qubo.ExactQUBOOptimizer()
result = optimizer.solve_qubo([[-2.0, 0.0], [0.0, -1.0]])
print(result.best_solution, result.best_value)

Named QUBO models and hardware-oriented precision reduction are available without another package:

from ascent import QuboBuilder, reduce_precision

builder = QuboBuilder()
builder.binary("x")
builder.binary("y")
problem = builder.add_linear_equality({"x": 1, "y": 1}, 1, penalty=5).build()
prepared = reduce_precision(problem.matrix, maximum_absolute=1.0, decimal_places=6)

The wheel is pure Python and universal across supported operating systems. Install the single ascent_science-0.7.0-py3-none-any.whl artifact; platform-specific wheel tags are not used.

After first activation, local solvers validate the cached license offline by default. Set ASCENT_LICENSE_VALIDATION=online when each solver construction must check revocation online.

One-command ant colony solver

After the SDK is installed and activated, run the built-in TSP example directly:

ascent-aco

Use a JSON or CSV distance matrix for your own problem:

ascent-aco --matrix distances.json --ants 50 --iterations 300 --start-city 0

JSON input may be a two-dimensional array or an object with a distance_matrix field. Add --json when the result will be consumed by another program.

One-command particle swarm solver

Run the default two-dimensional Sphere benchmark:

ascent-pso

Choose an included objective and tune the search space:

ascent-pso --objective rastrigin --dimensions 5 --lower -5.12 --upper 5.12 \
  --particles 50 --iterations 300

Available command-line objectives are sphere, rastrigin, and rosenbrock. Python callers pass any callable objective to ascent.combinatorial.PSOOptimizer.

CIM hardware execution

from ascent.cim import CIMOptimizer

optimizer = CIMOptimizer(
    task_name="production-qubo",
    endpoint="https://your-real-cim-gateway.example/v1",
)
task = optimizer.solve_qubo([[-2.0, 0.0], [0.0, -1.0]])
result = task.wait(timeout=600)
print(result.best_solution, result.best_value, result.cached)

Set ASCENT_CIM_ENDPOINT instead of passing endpoint in application code. The configured service must implement POST /tasks, GET /tasks/{id}, POST /tasks/{id}/cancel, and GET /tasks/{id}/logs.

First-use SDK activation stores an owner-only license at ~/.ascent/license.json. Hardware results are cached under ~/.ascent/cim-cache unless caching is disabled.