My inclination is that init = (args...) -> 2randn32(args...) or init = (args...) -> lmul!(2, randn32(args...)) isn't significantly more boilerplate and is a little more flexible in terms of supported operations. Perhaps documenting this pattern would be the way to go?

I feel the same way. A neat trick but seems like too much for something that is easily implemented as an anonymous or one line function.

Ok, I think I agree this is too much. Better to stay closer to Base... and if in doubt you should make the matrix yourself.