Nondeterminism even when setting all seeds, 0 workers, and cudnn.deterministic

[Naman-ntc (Naman Jain)]

@ptrblck

Basically, even if you set torch.backends.cudnn.deterministic = True , torch.backends.cudnn.benchmark = True might select different deterministic algorithms, which would yield different results.
There is a current PR working on this.

Could you disable benchmarking and try it again?

Hi, I am also facing a similar issue

 torch.backends.cudnn.deterministic=True
 torch.backends.cudnn.benchmark=False

I do just after the imports, but I am still getting non deterministic behaviour. I have also tried torch.backends.cudnn.enabled=False as I read that cunn modules also provide deterministic behaviour.

I have shuffle turned off, model on eval mode, and backpropogating for the gradients of input image (for performing adversarial attacks), so I believe protocol is fulfilled ?
Link to thread I created (which further contain links to questions I created)