To make better use of our shared resources, please follow these practices:
Kill the Jupyter Notebook kernel after finishing your experiment
Either write a small cleanup script at the end of your notebook or manually release the GPU.
Kill the Jupyter Notebook process completely once done
This ensures no background processes keep consuming GPU memory.
Use dedicated GPUs where possible
On cseproj158, multiple GPUs are available. Please run your experiments on a specific GPU by setting:
CUDA_VISIBLE_DEVICES=<gpu_id> in your environment, and
assigning models to a device with torch.to("cuda:0") (or whichever GPU you want to use).
Please do not kill someone else’s process
Avoid using sudo kill or pkill to terminate another person’s job, as this may disrupt important experiments. If you notice issues, kindly inform the group instead.
Tip: Convert notebooks to scripts for cleaner execution
You can convert your Jupyter Notebook into a Python script (.py) and run it directly on the server. This way, once the script finishes, all resources (including GPU memory) are automatically released.
By following these steps, we can ensure smoother GPU usage and avoid conflicts.