It's weird when I use distributed code

[Feywell]

I try to change code to use torch.ditributed. But, It will done without errors in

SEAN/models/networks/normalization.py

Line 119 in 04c7536

added_noise = (torch.randn(x.shape[0], x.shape[3], x.shape[2], 1).cuda() * self.noise_var).transpose(1, 3)

.
I found it will done with torch.randn(x.shape[0], x.shape[3], x.shape[2], 1).cuda() . It is weird.