Nabla

Records scalar functions once and evaluates them repeatedly with new inputs.

try (var evaluator = Nabla.model(rec -> {
    SDouble x = rec.input("x", 2.0);
    rec.output(x.mul(x));
}).fp64().gradients().on("cpu").build()) {
    var result = evaluator.value().with("x", 3.0).scenarios(1).run();
    System.out.println(result.value());          // 9.0
    System.out.println(result.gradient("x"));    // 6.0
}

Building an evaluator records and compiles the tape. Request inputs are kernel arguments, so subsequent evaluations reuse the compiled function.