Clarity on nondeterminism with torch.nn.functional.pad

Hello, I’d like to use torch.nn.functional.pad in the design of a neural network that will be trained on a GPU. I noticed the note on the docs page for pad:

When using the CUDA backend, this operation may induce nondeterministic behaviour in its backward pass that is not easily switched off. Please see the notes on Reproducibility for background.

However, when I checked the Reproducibility page, I wasn’t sure which—if any—of the sections applied to pad. There are several particular notes about cuDNN for convolutional operations, and I suspect that many CNN designs leverage pad, but my model does not use convolutions. How can I determine if the nondeterminism warning is relevant to me?