CUDA Setup failed despite GPU being available. Please run the following command to get more information: python -m bitsandbytes about bitsandbytes

[s-ravi18]

Same error I am also getting, can anyone suggest a way for this?

[johnga1995]

I had this issue while trying to get Kohya_ss setup and managed to solve it.

My system has cuda toolkit version 12.2.
NVIDIA-SMI 535.161.07 Driver Version: 535.161.07 CUDA Version: 12.2. If you have a lower driver version (i.e. ~320 with cuda 11.8) it should work just by installing bitsandbytes with pip. But I could not downgrade my driver version and the default installation of bitsandbytes does not seem to work with cuda 12.2.

So I managed to fix it with the help of #551 (comment)

I did have to do a few extra steps.

Basically made sure to uninstall bitsandbytes and bitsandbytes_windows from my venv environment and build it from source (as shown in the link above).

Since I had cuda 12.2, I had to specify the cuda version in the build process (refer to link) as CUDA_VERSION=122 make cuda12x

With this i generated a file libbitsandbytes_cuda122.so. This file I had to copy it to the venv python library path for bitsandbytes .

Also, on the venv, I had to set environment variables with commands export BNB_CUDA_VERSION=122 and export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda-12.2/targets/x86_64-linux/lib/stubs (Note that this is because I have cuda 12.2)

Finallly after running python -m bitsandbytes, I did not get any more errors and got a Installation was successful!.

[deep-pipeline]

@tejarao1156 and @s-ravi18 did the above advice help you? if so please say (or clarify any additions) so that this issue could be closed.