Run notebooks and jobs on UNC Longleaf

Complete repository and environment setup first. Use your own UNC username and the path to your research checkout in the commands and forms below.

Open OnDemand

Open UNC Open OnDemand and start a Jupyter session with a GPU request. Set the Jupyter startup directory to your research repository root, for example /nas/longleaf/home/YOUR_ONYEN/influpaint, and select the Python (diffusion_torch) kernel.

The early experiments used these additional job submission arguments:

Text Only

--mem=32gb -p volta-gpu --qos=gpu_access --gres=gpu:1

These are historical resource settings. The research notes report that a 16 GB Volta was insufficient for later runs and that Open OnDemand sessions could receive a small A100 MIG slice. Request enough GPU memory for the intended workload; the paper training launcher uses l40-gpu. Check the resources allocated to your session before starting training or inpainting.

Jupyter on a compute node

For a terminal-managed session, request a GPU node, then start Jupyter from the research repository root in your environment:

Bash

srun --ntasks=1 --cpus-per-task=4 --mem=32G --time=18:00:00 \
  --partition=l40-gpu --gres=gpu:1 --qos=gpu_access --pty bash
conda activate diffusion_torch
cd /nas/longleaf/home/YOUR_ONYEN/influpaint
hostname
jupyter lab --no-browser --ip=127.0.0.1 --port=8888

Wait for Slurm to allocate the node. On your local machine, replace COMPUTE_NODE with the hostname printed above and create the SSH tunnel through the login node:

Bash

ssh -N -J YOUR_ONYEN@longleaf.unc.edu \
  -L 8888:127.0.0.1:8888 YOUR_ONYEN@COMPUTE_NODE

Open the local Jupyter URL printed by the server, including its token. To use a password, run jupyter server password once in the environment before starting the server. The earlier README referred to create_notebook_password.sh and runjupyter.sh; those launch scripts are not present under those names in this checkout.

The original UNC-IDD patron-node example used partition jlessler with one GPU for 18 hours. The recorded node specifications were 512 GB RAM, 56 physical CPU cores (112 with hyperthreading), and four NVIDIA L40 GPUs with 48 GB each. Use that partition only with access to the allocation.

Paper batch jobs

The current launchers and command list are in main_training/. They contain the original author's Python environment path and Longleaf resource requests. Update those settings for your account and submit from the research repository root. Follow training, forecast-job preparation, and inpainting for the commands and required inputs.