GitLab

`torch.randn` requires `generator` and `latents` on the same device, while the wrapped function `randn_tensor` does not have this issue.