Minimize L as much as possible (this reduces mirroring accuracy)
Minimize W as much as possible (Pay attention to headroom)

Avoid a 1:9 ratio and use multiple stages, i.e. 1:3:9

Thank you for your kindness. You are right.As you said,W,L must be small as much as possible.When I change the W and L(randomly but of course be small),the ac response is much more better(simulation graph pasted below with R2/plus input signal;R0/plus,R1/plus  output signals).It is 10:10 ratio.
However,the input current is also pull down.I have to  tradeoff.Normally,input current can reach 20.8uA without any load.
The output signal,R1/plus,has a little ripple(the zoom in graph of R1/plus pasted below).
Besides,loose-electron,would you like to explain how to get the bandwidth of my design in detail? I can't quite follow you.Thank you again.

You need to explain a lot more clearly the specifics of what you are trying to do.

THere are active amplifier methods of driving gate bias to create a wider BW structure, but then its not a "current source" it is a common gate amplifier.

A "current source" is generally a fixed bias structure.