That's because octave is using doubles. You can do the exact same thing in PyTorch by passing in dtype=torch.float64 into torch.randn.