What impact will max_txt_len have on training?

[ZhengQinLai]

Thank you for your outstanding work! In the second stage of training, I commented out max_txt_len and the memory usage decreased by 10G (from 60G to 50G, A100). I want to know what impact max_txt_len will have on training. It seems that this option was not available in the first stage of pre-training. Thank you!