search

hustvl / Vim

[ICML 2024] Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model
Apache License 2.0
2.56k stars 160 forks source link

Incompatible to CU118 when installing causal-conv1d and mamba #21

Open wangf3014 opened 4 months ago

wangf3014 commented 4 months ago

Thanks a lot to the authors for the wonderful work. I have some quick questions about the repo.

When installing causal-conv1d and mamba, I encountered an error saying "The detected CUDA version (12.3) mismatches the version that was used to compile PyTorch (11.8). Please make sure to use the same CUDA versions.". Do you know how to address it?

I also tried several different torch and cuda versions (eg tc211+cu121), it seems that sometimes it can be successfully installed but the model training become quite slow. Did you find similar issues about that, or does this repo have a strict constraint of the torch and cuda versions?

pamia commented 3 months ago

Thanks a lot to the authors for the wonderful work. I have some quick questions about the repo.

When installing causal-conv1d and mamba, I encountered an error saying "The detected CUDA version (12.3) mismatches the version that was used to compile PyTorch (11.8). Please make sure to use the same CUDA versions.". Do you know how to address it?

I also tried several different torch and cuda versions (eg tc211+cu121), it seems that sometimes it can be successfully installed but the model training become quite slow. Did you find similar issues about that, or does this repo have a strict constraint of the torch and cuda versions?

I met the same problem with you (cuda version is 12.2 + cu11.8),do you successfully train the model with CUDA version (12.3)+ cu121

Magicboomliu commented 3 months ago

Me neither, seems this repo has a very strict constraint that the CUDA Version Must be 11.8.

laisimiao commented 3 months ago

Me neither, seems this repo has a very strict constraint that the CUDA Version Must be 11.8.

I don't think so. If you install caual-conv1d and mamba whthin this repo's directory and don't use compiled whl to install, I worked on cuda11.7.

MichaelWangGo commented 2 months ago

Same here, looks like only using their provided causal-conv1d and mamba can train the model, so it needs to meet very strict requirement of cuda version

lebron-2016 commented 2 months ago

Me neither, seems this repo has a very strict constraint that the CUDA Version Must be 11.8.

I don't think so. If you install caual-conv1d and mamba whthin this repo's directory and don't use compiled whl to install, I worked on cuda11.7.

Hi, did you install the mamba library using the setup.py file in the corresponding folder? What torch version are you using under cuda11.7?