Variable input sequence length

[ruoyxue]

Dear authors, This is an amazing work! I'm working with variable sequence lengths of video data. In one batch, there could be several videos with different frame numbers, and they will be padded to the same length. When I use transformer, I use attention masks to solve the problem of variable input lengths, but I do not see a similar mask in the Mamba forward function. Is there any solutions for dealing with variable lengths in a batch when using Mamba? Thanks!