CUDA kernel compilation failed when running start_linux.sh

[DIGist]

Running start_linux.sh, right before the model download options.

RuntimeError: The detected CUDA version (12.1) mismatches the version that was used to compile PyTorch (11.7). Please make sure to use the same CUDA versions.

CUDA kernel compilation failed.

Unclear on how to resolve.

[jllllll]

Try prepending python webui.py on line 56 with CUDA_PATH="$INSTALL_ENV_DIR".

Not a frequent Linux user, so not too sure on the syntax. I assume CUDA_PATH="$INSTALL_ENV_DIR" python webui.py Is the correct way to set the CUDA_PATH environment variable.

[xNul]

Yeah it looks like there could be another version of CUDA on your system and the installer is using it for some reason. CUDA 11 should have been installed, not 12.1. If @jllllll's answer doesn't work, maybe this will work: "$INSTALL_ENV_DIR/bin/python" webui.py.

[jllllll]

@xNul On Windows, I use the CUDA_PATH env variable to set which cuda toolkit I want to use when compiling. I have 7 versions of it installed. Some Python packages, such as bitsandbytes, search multiple other variables. On Linux, I believe it may be necessary to also set LD_LIBRARY_PATH="${INSTALL_ENV_DIR}/lib64:${LD_LIBRARY_PATH}".

[xNul]

I tried to reproduce by installing CUDA-enabled pytorch system-wide and then installing with the one-click installer but failed. It might have something unique to WSL, but I don't have a Linux computer connected to an Nvidia GPU to test on.

@jllllll gotcha and yeah you might be right

[xNul]

@DIGist Could you open up cmd_linux.sh, use the command set, and send the log? It will help me look through your environment for any misconfigured variables which may be causing this.

[github-actions[bot]]

This issue has been closed due to inactivity for 30 days. If you believe it is still relevant, please leave a comment below.