I hope that torch.randn_like can take generator=RNG argument as other random generation functions.

init_params = [torch.randn_like(param.data, generator=rng) for param in nn.parameters()]

I know that the equivalence

init_params = [torch.randn(param.size(), generator=rng).to(param.data) for param in nn.parameters()]

can work.
But for consistency, I still hope that all random related functions can take generator as an argument.