hotfinda / VideoMambaPro

Improving Mamaba performance on Video Understanding task
Apache License 2.0
28 stars 5 forks source link

z still being used #2

Open andypinxinliu opened 2 months ago

andypinxinliu commented 2 months ago

This is a great work. In the appendix, I see that no z in the mamba pro module but in the code, in projection will still generate z. Am I understanding it correctly?

Amshaker commented 1 month ago

Hi @andypinxinliu ,

Are you able to match the paper implementation with the current codebase?

Thanks! Abdelrahman.

hotfinda commented 1 month ago

In the code of VideoMambaPro, there is indeed a part about z. Mainly, we inherited most of the code of VideoMamba to facilitate writing. However, that of z is commented out based on a 'flag' setting.

rolson24 commented 1 month ago

Where is that z commented out? It seems to me like the code never runs your modified selective scan function. Did you change the cuda kernels for selective scan to also implement your changes?

hotfinda commented 3 weeks ago

Hi we didn't use the selective_scan_cuda module, instead, we write the function to replace the selective_scan_cuda module. More precisely, e.g., we use the function mamba_inner_ref to replace the forward function in class MambaInnerFn.