Last updated 2026-05-07

Containerized Dev Pods

Outside the platform processes, the stax repo also publishes a set of container images for remote development. They are not required by the runner or dashboard — they are a separate offering for situations where you want a beefier remote box (typically RunPod) that you SSH into from your laptop.

Images

Tag	Image	Purpose
`base` / `latest`	`ghcr.io/elloloop/stax:base`	General development. No CUDA. Light.
`inference`	`ghcr.io/elloloop/stax:inference`	Run LLMs. CUDA 12.4, PyTorch, vLLM, AutoGPTQ, AutoAWQ.
`train`	`ghcr.io/elloloop/stax:train`	Train models. CUDA 12.8, PyTorch 2.9, autoresearch, wandb, tensorboard.

Connect

The repo ships connect.sh, an SSH helper that handles initial setup (key exchange, hostname registration in ~/.ssh/config as stax):

./connect.sh <pod-ip> <port> --setup
ssh stax

Inference image

vllm serve Qwen/Qwen2.5-72B-Instruct-AWQ --host 0.0.0.0 --port 8000

Mount /models to persist downloads across pod restarts. The image bundles PyTorch (CUDA 12.4), vLLM, transformers, accelerate, bitsandbytes, AutoGPTQ, AutoAWQ, sentencepiece, tiktoken.

Train image

cd /workspace/autoresearch
uv run prepare.py    # one-time data prep
uv run train.py      # run a training experiment

Includes CUDA 12.8, PyTorch 2.9.1, uv, autoresearch, transformers, datasets, wandb, tensorboard.

Common tooling (all images)

Category	Tools
Editor	vim, neovim
Shell	bash, zsh, tmux, screen
Search	ripgrep, fd, fzf, bat
Git	git, git-lfs, GitHub CLI
Network	curl, wget, mosh, socat, rsync
Build	build-essential, cmake, python3, pip, Node.js 22
System	htop, lsof, jq, tree

Environment variables (recognized at startup)

Variable	Description
`PUBLIC_KEY`	SSH public key (RunPod standard).
`AUTHORIZED_KEYS`	Additional SSH keys.
`GITHUB_TOKEN`	Authenticates the bundled `gh` CLI.
`GIT_USER_NAME`	Git author name.
`GIT_USER_EMAIL`	Git author email.
`HF_TOKEN`	Hugging Face token (for gated models).
`STARTUP_SCRIPT`	Commands to run on container start.

Building locally

docker build -t stax:base       base/
docker build -t stax:inference  inference/
docker build -t stax:train      train/

Running the local platform inside a pod

The CLI, runner, and dashboard work the same inside the container as on your laptop. SSH in and run:

ssh stax
curl -fsSL https://raw.githubusercontent.com/elloloop/stax/main/install.sh | sh
export PATH=$HOME/.local/bin:$PATH

stax runner start --workspace /workspace/your-repo
# Tunnel port 4800 back to your laptop (e.g. via VS Code Remote, ssh -L, or RunPod's port forward)