Fix Backward loop issue on pytorch 2.5

[ybh1998]

When I tried to back propagate on PyTorch 2.5.1, the program stuck at loss.backward(). After some investigation, I found the following issue:

When I add a breakpoint at

torchinterp1d/torchinterp1d/interp1d.py

Line 155 in 4e7533d

gradients = torch.autograd.grad(

The grad_fn of ctx.saved_tensors[0] is <Interp1dBackward>, causing a loop backward call.

When I add a breakpoint at

torchinterp1d/torchinterp1d/interp1d.py

Line 149 in 4e7533d

ctx.save_for_backward(ynew, *saved_inputs)

The grad_fn of ynew is <AddBackward0>, which means the grad_fn of ctx.saved_tensors[0] is touched between forward and backward. I tried to add .detach() after calling interp1d, but the results were the same.

I suspect that the PyTorch auto_grad engine modifies the grad_fn of ynew in-place after some version. Adding a detach() in the forward function resolves this problem.

[Nikhil0504]

@aliutkus could you please push this since this was the error I was finding as well!