Is there some way to make the SA_module and FP_module deterministic

[Jianghanxiao]

Hi,
I use the SAmodule and FPmodule in my network, I try setting all random seeds for torch, numpy and random. However, the results are still not deterministic. Then I check the model, I find that with the same input, the SAmodule will give different outputs. Is there some way to make the SAmodule deterministic?

I think the potential reason for this is that it uses some functions implemented in C, then the random seed cannot be controlled. Do you have some idea on this?

[kefusong]

I found out the atomicAdd() function in the .cu file is not fixed in order when called, resulting in a numerical error in the sum of floating points as a result of gradient calculations.
I also found that many networks are based on this package, and how do they deal with this non-determinism?
If determinism is not guaranteed, ablation experiments can be cumbersome.
Thank you!