Pod Inheritance¶

Standard OCI Containers

This hierarchy shows build inheritance, not deployment requirements. All workloads are published to ghcr.io/atrawog/bazzite-ai-pod-*:stable and run anywhere containers are supported.

Inheritance Tree¶

graph TD
    base[pod-base<br/>~7GB<br/>Fedora 43 + Dev Tools]

    base --> nvidia[pod-nvidia<br/>~8GB<br/>CUDA + cuDNN + TensorRT]
    base --> devops[pod-devops<br/>~10GB<br/>AWS + gcloud + kubectl]
    base --> runner[pod-githubrunner<br/>~8GB<br/>GitHub Actions Runner]

    nvidia --> python[pod-nvidia-python<br/>~14GB<br/>PyTorch ML via pixi]
    nvidia --> ollama[pod-ollama<br/>~11GB<br/>LLM Inference Server]

    python --> jupyter[pod-jupyter<br/>~17GB<br/>JupyterLab Server]
    python --> comfyui[pod-comfyui<br/>~26GB<br/>AI Image Generation]

    style python fill:#4CAF50,color:#fff
    style jupyter fill:#4CAF50,color:#fff
    style devops fill:#4CAF50,color:#fff
    style ollama fill:#4CAF50,color:#fff
    style comfyui fill:#4CAF50,color:#fff

Green nodes = Core workloads (recommended for most workflows)

Layer Structure¶

Each workload inherits tools from its parent, adding specialized functionality:

Layer 1: Base Foundation¶

pod-base (~7GB) - Clean Fedora 43 with development essentials

Build toolchain (gcc, make, cmake, ninja)
Language runtimes (Python 3.13, Node.js 23+, Go, Rust)
VS Code, Docker CLI, Podman
kubectl, Helm, Claude Code
Modern shell tools (fzf, ripgrep, bat, eza)

Layer 2: Specializations¶

From base:

Pod	Adds	Use Case
nvidia	CUDA 13.0, cuDNN, TensorRT	Custom GPU setups
devops	AWS, gcloud, Firebase, Grafana tools	Cloud infrastructure
githubrunner	GitHub Actions runner agent	CI/CD pipelines

Layer 3: ML/AI & LLM¶

From nvidia:

Pod	Adds	Use Case
nvidia-python	PyTorch, torchvision, torchaudio via pixi	ML/AI development
ollama	Ollama LLM server, model management	LLM inference

Layer 4: Interactive¶

From nvidia-python:

Pod	Adds	Use Case
jupyter	JupyterLab server	Interactive notebooks
comfyui	ComfyUI, Stable Diffusion	AI image generation

Image Registry¶

All workloads are published to GitHub Container Registry:

ghcr.io/atrawog/bazzite-ai-pod-<variant>:<tag>

Available Tags¶

Tag	Description
`stable`	Production-ready release
`latest`	Most recent build
`<version>`	Specific version (e.g., `1.0.0`)

Pull Examples¶

# Docker
docker pull ghcr.io/atrawog/bazzite-ai-pod-nvidia-python:stable

# Podman
podman pull ghcr.io/atrawog/bazzite-ai-pod-nvidia-python:stable

# Apptainer (converts to SIF format)
apptainer pull docker://ghcr.io/atrawog/bazzite-ai-pod-nvidia-python:stable

Build System¶

Workloads are built using a unified buildcache for efficient multi-variant builds:

pods/
├── base/Containerfile           # Base layer
├── nvidia/Containerfile         # CUDA layer
├── nvidia-python/Containerfile  # PyTorch layer
├── jupyter/Containerfile        # JupyterLab layer
├── comfyui/Containerfile        # AI image generation layer
├── ollama/Containerfile         # LLM inference layer
├── devops/Containerfile         # DevOps tools
└── githubrunner/Containerfile   # CI/CD runner

Build Commands¶

# Build specific workload
just pod build nvidia-python

# Build all workloads
just pod build all

# Push to registry
just pod push nvidia-python

Common Base Components¶

All workloads include (inherited from base):

Languages & Runtimes¶

Language	Version
Python	3.13
Node.js	23+
Go	Latest
Rust	Latest
.NET	8.0
PHP	Latest
Java	OpenJDK
Ruby	Latest

Development Tools¶

Category	Tools
Build	gcc, g++, make, cmake, ninja, meson
Version Control	git, gh CLI
Containers	Docker CLI, Podman
Kubernetes	kubectl, Helm
Editor	VS Code (code-server)
AI	Claude Code CLI

Shell Environment¶

Tool	Purpose
Starship	Modern shell prompt
fzf	Fuzzy finder
zoxide	Smart directory navigation
ripgrep	Fast search
bat	Better cat
eza	Better ls

Container User¶

All workloads run as user jovian (UID 1000) by default:

Username: jovian
UID: 1000
Home: /home/jovian
Workspace: /workspace (mounted from host)