inference in bidirectional mamba

Hello everyone, I wanted to ask if anyone has used Bidirectional Mamba during inference. It was introduced in the VIM paper, but as I understand, it's currently supporting full sequence not token by token. 

Has anyone tried implementing it for inference, or would specific modifications be needed for this? Any advice or insights would be greatly appreciated!

Thanks in advance!