Backward behaviour of scatter_max

[Elias Ramzi]

Hi,

thank you for the great repo!

I had a question regarding the backward behaviour of scatter_max.
Is it similar to the one of torch.max, i.e. only one index gets the gradient. Or like the one of torch.amax, i.e. in case of equality all indexes get gradient?

This difference is explained for torch here: https://pytorch.org/docs/stable/generated/torch.amax.html.

Thank you for your help!

[Matthias Fey]

Only one index gets the gradient.

[Elias Ramzi]

Thanks!