No, its not necessary to take the same lag length while applying the ARDL. In the econometric section of a research paper, the same lag length is shown to give a symmetric touch. However, practically the lag length may differ.

plz also tell me that after ardl, whether granger causalty is applied on the basis of var or vecm? 

With ARDL, VECM is better if you are interested in short- and long-run causalities. I suggest to follow the link:

http://davegiles.blogspot.com/2011/10/var-or-vecm-when-testing-for-granger.html

You can follow Dave's guides regrading your earlier question on Lag selections.  For VAR, an extra lag (correction of serial correlation) is included. Also here is an example:

If the unit root results suggest that the maximum order of integration is 1 i.e. dmax = 1, and the lags used for the ARDL estimation is 5 (l = 5). Hence, the maximum lags that can be used to carry out the non-causality test is dmax + l = 6.

Best,

Jawad 

Qurra, ARDL- atuo-regressive distributed lag as the name implies that the user can allocate the lag length in the work but not necessarily same lag length. As for the "granger causality test" you have to rethink it to form may be "Qurra causality test" before you can apply same lag length or not. I mean granger developed his model on specific lag lengths so anything outside that is no more granger same apply to ARDL. However, there are possibilities to what you are looking for. 

use either aic or bic to determine the lag length