In theory samples from an N(0, 1) distribution can get arbitrarily large, that is, outside of the range of whatever floating point type you’re using. Is it possible that the output of torch.randn can be Inf or NaN?
If it is guaranteed to be finite, is the range (given the type) known or even well defined?
I couldn’t find anything about that in the documentation:
https://pytorch.org/docs/stable/generated/torch.randn.html#torch.randn