Configuring a deep learning environment is a hurdle many struggle to overcome. However, with large models, troubleshooting and pinpointing issues can be significantly faster.

I spent some time adapting an older version of PaddlePaddle and finally got it working. Here, I’ll share an article documenting the process.

In Docker images, many CUDA 11-based images fail to run in a CUDA 12 environment. The exact reason isn’t entirely clear to me. In such cases, you can opt for a CUDA version that matches the major release.

To avoid affecting the environments of others on the server, do not update the NVIDIA driver. Instead, install your own CUDA version and modify the environment variables to change the system’s CUDA.

1
2
3
4

# CUDA_VERSION=11.7
export CUDA_HOME="/usr/local/cuda-$CUDA_VERSION"
export LD_LIBRARY_PATH="$CUDA_HOME/lib64:$LD_LIBRARY_PATH"
export PATH=$CUDA_HOME/bin:$PATH

Apply this environment variable, then check nvidia-smi to see the version change.