Confusing warning emitted from Tensor::scatter_reduce.

[iago-lito]

The following program:

use tch::{Device, Kind, Tensor};

fn main() {
    // Columns of a conceptual (6, 3) sparse matrix.
    let a = Tensor::from_slice(&[7.3, 5.5, 6.1, 3.3]).set_requires_grad(true);
    let b = Tensor::from_slice(&[8.8, 4.6]).set_requires_grad(true);
    let c = Tensor::from_slice(&[4.9, 8.9, 6.7]).set_requires_grad(true);
    // Indices of every columns value.
    let ia = Tensor::from_slice(&[0, 2, 3, 5]);
    let ib = Tensor::from_slice(&[1, 3]);
    let ic = Tensor::from_slice(&[1, 2, 4]);
    let src = Tensor::cat(&[a, b, c], 0);
    let idx = Tensor::cat(&[ia, ib, ic], 0);
    let rowsums =
        Tensor::zeros(6, (Kind::Double, Device::Cpu)).scatter_reduce(0, &idx, &src, "sum");
    println!("rowsums: {rowsums}"); // [7.3, 13.7, 14.4, 10.7, 6.7, 3.3].
} 

yields

Warning: The reduce argument of torch.scatter with Tensor src is deprecated and will be removed in a future PyTorch release. Use torch.scatter_reduce instead for more reduction options. (function operator())

As it turns out, adding the following lines makes it panic:

    let total = rowsums.sum(Kind::Double);
    total.backward();

called `Result::unwrap()` on an `Err` value: Torch("derivative for aten::scatter is not implemented")

I eventually figured that the fix is to replace .scatter_reduce(0, &idx, &src, "sum") by .reduce_add(0, &idx, &src), but the whole process has been rather confusing. What is the actual meaning of the warning I had? Is it actually related to the later panic?