February 23, 2024
You need UVa VPN to access Rivanna. The standard one “UVa Anywhere” should work. Then, login to the server using the following command
ssh [computing-ID]@portal01.cs.virginia.edu
with the CS password for your computing ID.
To avoid any unexpected interruption on the configuration process, we can use a tmux session
tmux new -s llmtuning
In case of any interruption, we can log back in using the following command and continue the configuration
tmux a -t llmtuning
Before running anything, you need to load some relevant modules as you did in the CS department servers. For Rivanna, the module names are slightly different
module load anaconda3-2023.09 cuda-toolkit-11.8.0 cudnn-8.4.1_cuda11.x
One way to figure out the cuda version is to use the following command after load the cuda and cudnn modules
nvcc -V
By default, the cuda version is 12.2.
Assume the virtual environment name is test and the Python version is 3.9, we use
conda create -n test python=3.9
If a conda environment has already been activated, such as