faeyon

Consider a simple model like this:

from torch import nn
from faeyon import faek

with faek:  # or faek.on() to enable process-wide
    model = nn.Linear(10, 5) >> nn.Linear(5, 3)

Importing faeyon does not change PyTorch's behavior. The nn.Module interception (faek) is opt-in: enable it explicitly with faek.on(), or scope it to the code that builds expressions with with faek:. Evaluating or materializing an already-built tree (data | expr, FaeModule, lower) does not require it.