UOA Cloud Server: Setup and Usage Guide

This guide walks through everything you need to start running deep learning workloads on the UOA compute cluster — from initial login to submitting GPU jobs and monitoring training.

0. First-Time Setup

When you first get access to your machine, point your data and home directories at the large /data/<UPI>/ partition so that datasets, checkpoints, and virtual environments don't fill up the small system disk.

1
2

export TMPDIR=/data/<UPI>/
export HOME=/data/<UPI>/

Tip: add these lines to your ~/.bashrc or ~/.zshrc so they persist across sessions.

1. SSH Login

1
2

ssh <UPI>0@foscsmlprd01.its.auckland.ac.nz
# foscsmlprd02 and foscsmlprd03 are also available

Replace <UPI> with your University of Auckland UPI (e.g., jdoe001).

2. Python Environment with uv

We use uv for fast, reproducible Python environment management.

Install uv

1

wget -qO- https://astral.sh/uv/install.sh | sh

Set Up an Environment

If the project has a pyproject.toml:

1

uv sync

uv creates the virtual environment and installs all dependencies in one step.

If the project only has a requirements.txt:

1
2

uv venv                              # create the virtual environment
uv pip install -r requirements.txt   # install dependencies

3. Submitting GPU Jobs with SLURM

To run training on a GPU, submit a shell script to the SLURM scheduler — don't run heavy workloads directly on the login node.

Example Job Script

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

#!/bin/bash

# ======== SLURM Job Configuration ========
#SBATCH --job-name=<job_name>             # [REQUIRED] Name shown in the job queue (e.g., "AAAI-training")
#SBATCH --time=00:30:00                   # [REQUIRED] Wall-clock limit (HH:MM:SS)
#SBATCH --open-mode=append                # Append to log files instead of overwriting
#SBATCH --output=<output_file>.log        # [RECOMMENDED] Standard output log
#SBATCH --error=<error_file>.log          # [RECOMMENDED] Standard error log
#SBATCH --gres=gpu:1                      # [REQUIRED] Number of GPUs (>1 needs Slack channel approval)

# ======== Job Execution ========

# Navigate to your project directory
cd /data/<your_username>/<project_path>

# Activate the virtual environment
source venv/bin/activate

# Run the training script
python3 train.py

Common SLURM Commands

4. Monitoring Your Workflow

Watch your job queue in real time:

1

watch squeue -u $USER

Tail a log file to follow output or errors live:

1

tail -f logs/xxx.txt

This is the fastest way to debug a running job without waiting for it to finish.

5. Downloading Output Files

For terminal users who want to pull results back to a local machine:

1

scp <UPI>0@foscsmlprd03.its.auckland.ac.nz:/data/<UPI>/LongLive/videos/interactive/rank0-0-0_lora.mp4 ~/Nexus/ai_build/new.mp4

The pattern is scp <user>@<host>:<remote_path> <local_path>.