Apptainer - HPC Container Management

Overview

The apptainer command manages Apptainer (formerly Singularity) containers for HPC-compatible workloads. It provides SIF image management with automatic GPU detection.

Key Concept: Apptainer is the HPC standard. Unlike Docker/Podman, containers run as the user (no root). SIF files are single-file images.

Quick Reference

Action	Command	Description
Build	ujust apptainer build DEF	Build SIF from definition file
Cache	ujust apptainer cache [clean|status]	Manage Apptainer cache
Exec	ujust apptainer exec IMAGE CMD	Execute specific command in container
Inspect	ujust apptainer inspect IMAGE	Show SIF file metadata
Pull	ujust apptainer pull IMAGE	Download container image to SIF file
Run	ujust apptainer run IMAGE	Run container with default command
Shell	ujust apptainer shell [-- CMD]	Open interactive shell in container

Parameters

Parameter	Long Flag	Short	Default	Description
action	(positional)	-	required	Action: pull, run, shell, exec, build, inspect, gpu, cache
image	--image	-i	""	SIF file path, image name, or DEF file
tag	--tag	-t	""	Image tag, output file, or cache subaction
cmd	(variadic)	-	""	Command to execute (use -- separator)

Pull Images

overthink Pod Images

# Pull nvidia-python (long form)
ujust apptainer pull --image=nvidia-python

# Pull with tag (long form)
ujust apptainer pull --image=nvidia-python --tag=testing

# Pull nvidia-python (short form)
ujust apptainer pull -i nvidia-python

# Pull with tag (short form)
ujust apptainer pull -i nvidia-python -t testing

# Pull jupyter
ujust apptainer pull --image=jupyter --tag=stable

External Images

# Docker Hub
ujust apptainer pull --image=docker://ubuntu:22.04

# NVIDIA NGC
ujust apptainer pull --image=docker://nvcr.io/nvidia/pytorch:latest

# Sylabs Cloud
ujust apptainer pull --image=library://sylabsed/examples/lolcow

Pull Output

Images are saved as SIF files:

~/.local/share/apptainer/overthink-pod-nvidia-python.sif

Run Containers

Run with Default Command

# Run nvidia-python (long form)
ujust apptainer run --image=nvidia-python

# Run nvidia-python (short form)
ujust apptainer run -i nvidia-python

# Run specific SIF file
ujust apptainer run --image=./my-container.sif

Run with Command

# Run Python in container (use -- separator for commands)
ujust apptainer run --image=nvidia-python -- python

# Run script
ujust apptainer run --image=nvidia-python -- python script.py

# Short form
ujust apptainer run -i nvidia-python -- python train.py

GPU Auto-Detection

GPU flags are auto-detected:

• NVIDIA: Adds --nv
• AMD: Adds --rocm

# GPU is automatically enabled
ujust apptainer run --image=nvidia-python -- python -c "import torch; print(torch.cuda.is_available())"

Interactive Shell

# Shell into container (long form)
ujust apptainer shell --image=nvidia-python

# Shell into container (short form)
ujust apptainer shell -i nvidia-python

# Now inside container
python --version
nvidia-smi
exit

Execute Commands

# Execute single command (use -- separator)
ujust apptainer exec --image=nvidia-python -- pip list

# Execute Python one-liner
ujust apptainer exec -i nvidia-python -- python -c 'print(1+1)'

Build from Definition

Definition File Example

Bootstrap: docker
From: ubuntu:22.04

%post
    apt-get update
    apt-get install -y python3 python3-pip

%runscript
    python3 "$@"

Build

# Build SIF from definition (image=DEF, tag=OUTPUT)
ujust apptainer build --image=mydef.def --tag=myimage.sif

# Build to default location
ujust apptainer build --image=mydef.def

# Short form
ujust apptainer build -i mydef.def -t myimage.sif

GPU Support

Test GPU

# Detect and test GPU
ujust apptainer gpu

GPU Flags

GPU	Flag	Auto-Detection
NVIDIA	--nv	Yes
AMD	--rocm	Yes
Intel	(none yet)	No

Manual GPU Override

# Direct apptainer command with GPU
apptainer run --nv nvidia-python.sif nvidia-smi

Cache Management

List Cache

# Long form
ujust apptainer cache --tag=list

# Or
ujust apptainer cache list

Clean Cache

# Long form
ujust apptainer cache --tag=clean

# Or
ujust apptainer cache clean

Cache is stored in ~/.apptainer/cache/.

Common Workflows

HPC Development

# Pull HPC-ready image
ujust apptainer pull --image=nvidia-python

# Test GPU
ujust apptainer gpu

# Development shell
ujust apptainer shell --image=nvidia-python

# Run production workload
ujust apptainer run --image=nvidia-python -- python train.py

Use NGC Images

# Pull NVIDIA PyTorch
ujust apptainer pull --image=docker://nvcr.io/nvidia/pytorch:23.10-py3

# Run training
ujust apptainer run --image=pytorch_23.10-py3.sif -- python train.py

Build Custom Image

# Create definition file
cat > myenv.def << 'EOF'
Bootstrap: docker
From: python:3.11

%post
    pip install numpy pandas scikit-learn

%runscript
    python "$@"
EOF

# Build
ujust apptainer build --image=myenv.def --tag=myenv.sif

# Test
ujust apptainer run --image=myenv.sif -- python -c "import numpy; print(numpy.__version__)"

Apptainer vs Docker/Podman

Feature	Apptainer	Docker/Podman
Root required	No	Sometimes
Single file	Yes (SIF)	No (layers)
HPC compatible	Yes	Limited
GPU support	--nv, --rocm	nvidia-docker
Security model	User namespace	Container namespace

Use Apptainer when:

• Running on HPC clusters
• Need single-file portability
• Can't run as root
• Need reproducibility

Troubleshooting

Pull Failed

Check:

# Test network
curl -I https://ghcr.io

# Check registry auth
apptainer remote list

Fix:

# Login to registry
apptainer remote login docker://ghcr.io

GPU Not Available

Check:

ujust apptainer gpu
nvidia-smi  # or rocm-smi

Fix:

# Ensure drivers installed
# For NVIDIA:
nvidia-smi
# For AMD:
rocm-smi

SIF File Corrupted

Fix:

# Remove and re-pull
rm ~/.local/share/apptainer/*.sif
ujust apptainer pull --image=nvidia-python

Cache Too Large

Check:

du -sh ~/.apptainer/cache/

Fix:

ujust apptainer cache --tag=clean

Cross-References

• Related Skills: pod (build OCI images), jupyter (uses containers)
• GPU Setup: ujust config gpu setup
• Apptainer Docs: https://apptainer.org/docs/

When to Use This Skill

Use when the user asks about:

• "apptainer", "singularity", "HPC container"
• "SIF file", "pull image", "build container"
• "apptainer GPU", "run with GPU"
• "HPC workload", "cluster container"