Fast Mamba kernels are not available

#16
by MohamedRashad - opened

I installed everything as instructed and still gives me this error, why ?

Same. Had to disable them in my finetune demo (but may be the cause of crazy loss jump…)

I am also facing same problem.

Ok found an indirect fix: import mamba-ssm before doing anything with the model. If there is an issue (for instance version conflict with another library) it will be immediately outputted and bring the possibility to correct it.

@Pclanglais Didn't change anything

@MohamedRashad when doing import mamba-ssm , do you have any error message?
Could come otherwise from causal-conv1d, a key dependency that may also run into compilation issues.

@Pclanglais No, it gives me no errors

I met same problem and found that causal-conv1d install cuda12 version via pip so I built from source

To investigate, you need to remove try-catch of import causal-conv1d not one of import mamba-ssm

Have you tried to install causal-conv1d and mamba from source?
pip instal git+https://github.com/Dao-AILab/causal-conv1d
pip instal git+https://github.com/state-spaces/mamba

Thanks @SpiridonSunRotator ! It worked for me.

Sign up or log in to comment