man advect/api/interop/torch
TORCH(3)Library CallsTORCH(3)
────────────────────────────────────────────────────────────────────────────────────────

PyTorch

The shared bridge contract and host-framework tutorial explain what crosses this boundary.

wrap 

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

Wrap a NumPy-backed callable as a first-order PyTorch operation.

Every tensor leaf in positional or keyword arguments with a NumPy floating or complex representation is an Advect input. Static configuration should be closed over by function. Values execute through NumPy on the host and outputs return to the inputs' common device. One PyTorch backward consumes the retained Advect pullback.

import numpy as np
import torch

from advect.interop.torch import wrap


def numpy_energy(value):
    return np.sum(np.sin(value) ** 2)


energy = wrap(numpy_energy)
x = torch.linspace(0, 1, 8, requires_grad=True)
energy(x).backward()
print(x.grad)

Inputs are copied through host NumPy. Outputs return to the common input device, and input gradients restore each input tensor's device and dtype. Eager torch.autograd is the supported boundary. One backward consumes the retained Advect Pullback; repeated backward over the same retained PyTorch graph is not supported. PyTorch dtypes without a NumPy representation, including bfloat16, Float8, and complex32, reject at the boundary.

$ █

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