gabszap/base-image

Personal maintained fork of lucidboxai/base-image, which itself is a modernized fork of ai-dock/base-image by @robballantyne. This is not the official image — it's a personal fork for personal use, with no warranty or support guarantees.

Fork lineage

This image inherits all the modernization work that lucidboxai applied to the upstream ai-dock base:

The table below is historical context for the inherited lucidboxai baseline, and the current gabszap v3 stack.

Component	Upstream ai-dock	lucidboxai	gabszap v3
Ubuntu	22.04	24.04	CPU: 22.04 / NVIDIA: 24.04
CUDA base	11.8 / 12.1	12.6.3-cudnn-runtime	13.0.3-cudnn-runtime
Python	3.10	3.12 (native, no PPA)	CPU: 3.10 / NVIDIA: 3.12 (native, no PPA)
Build toolchain	Go 1.22, xcaddy 0.4.2, Node 22.2	Go 1.26.3, xcaddy 0.4.5, Node 22.22.3	Go 1.26.3, xcaddy 0.4.5, Node 24.18.1, npm 12.0.2
FastAPI / uvicorn	hard-pinned ==0.103 / ==0.23	floor-pinned >=0.115 / >=0.34	floor-pinned >=0.115 / >=0.34
Targets	CPU / NVIDIA / AMD	CPU / NVIDIA / AMD	CPU / NVIDIA (AMD/ROCm removed)

Ubuntu 24.04 migration fixes (package renames, user cleanup, dropped unsupported packages) are inherited from the lucidboxai fork. See lucidboxai/base-image for the full compatibility reference.

PyTorch is not installed in this image.

CUDA 13 NVIDIA variants support GeForce RTX 50-series / Blackwell GPUs; support depends on a compatible host NVIDIA driver and NVIDIA Container Toolkit for GPU runtime.

Build tags:

• ghcr.io/gabszap/base-image:v3-cpu-22.04
• ghcr.io/gabszap/base-image:v3-cuda-13.0.3-base-24.04
• ghcr.io/gabszap/base-image:v3-cuda-13.0.3-cudnn-runtime-24.04 (recommended)
• ghcr.io/gabszap/base-image:v3-cuda-13.0.3-cudnn-devel-24.04

What I added

On top of everything lucidboxai already did, I threw in a few extra things:

• VS Code (code-server) pre-configured — Runs VS Code in the browser via supervisor. Pre-configured with: Copilot disabled, workspace trust off, telemetry off, custom icons and color theme installed.
• Dark mode for the portal — Full dark theme across the Service Portal, plus a migration to FastAPI's newer TemplateResponse API. Long overdue since only light mode existed before.
• Inline utilities — Several helper scripts that used to be symlinks are now real files with better error handling.
• OpenCode Integration — AI-powered coding assistant integrated into the dev environment. (run with opencode)
• Tmux — Terminal multiplexer for multiple sessions. See shortcuts guide.
• Superfile — A file manager TUI for browsing files in the terminal. See shortcuts guide.
• Neovim — A terminal-based text editor. See shortcuts guide.
• eza — A modern replacement for ls with a better terminal UI. See shortcuts guide.
• zoxide — A smarter cd with binary setup. See shortcuts guide.
• fzf — A fuzzy finder with optional shell CLI setup. See shortcuts guide.

Looking for shortcuts? For a quick reference on how to use the included tools (Tmux, Neovim, etc.), check out our Shortcut Guide.

Need help with icons? To properly render icons and glyphs in tmux and your prompt/status bars, please install and select a Nerd Font in your local terminal application. Without a Nerd Font, these symbols may appear as boxes or broken symbols. Popular examples include JetBrainsMono Nerd Font. You can browse and download fonts via nerdfonts.com or the official GitHub releases.

Documentation

Container behavior and environment variables shared with upstream are still documented in the ai-dock base wiki. lucidboxai-specific deviations from those docs are tracked in the lucidboxai repo's commit history; changes in this fork are listed above.

Building locally

git clone https://github.com/gabszap/base-image.git
cd base-image

# GPU build (default)
docker compose --env-file .env.gpu.example build
docker compose --env-file .env.gpu.example up -d --no-build

# CPU build
docker compose --env-file .env.cpu.example build
docker compose --env-file .env.cpu.example up -d --no-build

You may copy either example to .env for customization and replace placeholder credentials. Building the GPU image does NOT require a GPU, but GPU-accelerated Compose runtime requires a compatible NVIDIA driver, NVIDIA Container Toolkit, AND enabling/uncommenting the NVIDIA GPU reservation block in docker-compose.yaml. The image base, CUDA string, and tag are all set via docker-compose.yaml build args — override via env vars if you need a different variant.

Run with docker compose exec supervisor bash.

Credits

• lucidboxai/base-image — maintained by lucidboxai. Their Ubuntu 24.04 migration, CUDA 12.6 modernization, and Python 3.12 groundwork make up the bulk of the improvements over upstream.
• ai-dock/base-image — by @robballantyne. The original architecture, build scripts, Caddy + supervisor setup, and ai-dock common-base concept are the foundation everything in this lineage builds on.

This is a personal fork for personal use. It comes with no warranty or support guarantees — use at your own risk.