man advect/api/interop/autograd
AUTOGRAD(3)Library CallsAUTOGRAD(3)
────────────────────────────────────────────────────────────────────────────────────────

HIPS Autograd

This adapter carries an Advect pullback into HIPS Autograd through autograd.grad. See the shared bridge contract and host-framework tutorial for the boundary around the wrapped call.

wrap 

wrap(
    function: Callable[..., object],
) -> Callable[..., object]

Wrap a NumPy-backed callable as a first-order HIPS Autograd primitive.

Every NumPy floating or complex leaf in positional or keyword arguments is selected. The bridge translates between Autograd's complex-bilinear cotangents and Advect's real-adjoint convention. The exact forward linearization remains reusable for first-order host transforms; higher-order differentiation is rejected.

import autograd
import autograd.numpy as anp
import numpy as np

from advect.interop.autograd import wrap


energy = wrap(lambda value: np.sum(np.sin(value) ** 2))
gradient = autograd.grad(energy)(anp.linspace(0, 1, 8))
print(gradient)

The custom primitive retains the exact reusable Advect LinearMap from its forward call. First-order transforms such as autograd.jacobian may apply its pullback more than once; releasing the host VJP releases the retained linearization. A nested Autograd trace raises an explicit higher-order error.

$ █

[1:docs] [2:playground] $ man advect/api/interop/autograd