[Help needed] How to use max_norm in embedding

And do we normally use that in current SOTA models (any kind of LSTMs)? I have heard someone says that to be set as 10^3 to 10^4, is that correct? Thanks!