It was originally normalized on the go:

for (size_t i = 0; i < outer_dims; ++i)
{
  // ...
  for (size_t j = 0 ; j < indices.total(); ++j)
  {
    const size_t index = (static_cast<int>(idx[j]) + inp.size[axis]) % inp.size[axis];
    // ..
  }
}

I think in onnx importer we should load what it is without any extra operation on constant input & attributes. Operations like normalization should be done in the layer constructor or somewhere. We need the original information when it comes to build operators for other backend, like CANN and TIM-VX, who have operator primitives aligned with ONNX, TF ...

"On the go" normalization is not efficient as performs it multiple times (especially through division).
There is change to perform normalization once (per forward() call) before the main loop.

It was done because the elements of index tensor can be negative per spec and non-constant at the same time. We cannot perform normalization ahead of time since we don't have the data.