When I run mamba2 : ImportError: libcudart.so.11.0: cannot open shared object file: No such file or directory

[Rruixu]

import selective_scan_cuda ImportError: libcudart.so.11.0: cannot open shared object file: No such file or directory

my cuda : 12.1

Is cuda 12.+ dont match mamba2 ?

[JerayuT]

@Rruixu I also ran into the same problem. Did you solve end up figuring it out?

[RohollahHS]

@Rruixu I also ran into the same problem. Did you solve end up figuring it out?

Hi. Did you solve this issue?

[JerayuT]

I'm not sure if this will help, but I ran into this issue when I tried after installing triton separately in a conda environment to make it match their triton version. I fixed it but just creating a new environment, using the latest version of torch and installing causal-conv1d followed by mamba-ssm and it seemed to fix the issue.

[rozariwang]

I'm not sure if this will help, but I ran into this issue when I tried after installing triton separately in a conda environment to make it match their triton version. I fixed it but just creating a new environment, using the latest version of torch and installing causal-conv1d followed by mamba-ssm and it seemed to fix the issue.

Which cuda version are you using?

[JerayuT]

I'm not sure if this will help, but I ran into this issue when I tried after installing triton separately in a conda environment to make it match their triton version. I fixed it but just creating a new environment, using the latest version of torch and installing causal-conv1d followed by mamba-ssm and it seemed to fix the issue.

Which cuda version are you using?

I'm using version 12.4

[rozariwang]

Thanks! I am using 12.4 as well, and I followed the same steps as you did, but I keep facing the same issue. Are you installing mamba-ssm with pip or from source? What is your setuptools version?

[JerayuT]

Thanks! I am using 12.4 as well, and I followed the same steps as you did, but I keep facing the same issue. Are you installing mamba-ssm with pip or from source? What is your setuptools version?

I'm using setuptools 0.9.8 and installing mamba-ssm using pip. You might have to delete your site-packages? Here are the exact steps I took to fix this issue.

Deleted all python site-packages (From what I saw creating a new conda environment will still refer to packages in here which is why it couldn't find the correct libcudart file so I opted to delete everything and reinstall them) Created a conda environment (I used python=3.9.5) conda installed cuda-nvcc pip install torch, causal-conv1d >= 1.4.0 and then mamba-ssm using --no-cache-dir.

After that I could run mamba 2. Hope this helps!

[rozariwang]

@JerayuT Thank you so much! I'm using Docker, and it turned out that I had to switch the base image. basically what you did as well! Thanks again

[cycDL]

@rozariwang Hello, have you finally solved this problem (import selective _ scan _ cudayimporterror: librcudart.so.11.0: cannot open shared object file: no such file or directory)?

[rozariwang]

Hi @cycDL, yes, I did by switching the base image to nvidia/cuda:12.4.0-devel-ubuntu22.04. My context is a bit different though. Issue #567 for your reference.