Why is the implementation of Mamba so slow?

[baisonmg]

I haven't run the official version of Mamba, but I've run your implementation, and it seems that the training speed of this model is much slower than that of the Transformer.

[aryanmangal769]

I am also facing the similar issue. Are you able to find the reason for it?

[lingxitong]

that is mainly because there is something wrong which lead the loss.backward so slow

[johnma2006]

It's not meant to be fast! This repo is mostly only for educational purposes. I would suggest using the official repo to do any training: https://github.com/state-spaces/mamba

[D-Walter]

probably because mamba-ssm designed a GPU-adapted scanning operation based on C.

[mengdeyu539]

I think the problem is embeding layer. D-layer is 50280, lol.