Linux contributor installation

The public channel includes Linux x86_64 and ARM64 binaries. Run these commands in a Linux terminal on the contributing machine, under your normal user account.

1. Identify the machine

uname -m

x86_64 selects the x64 release; aarch64/arm64 selects ARM64. Automatic managed vLLM provisioning is specifically for detected ARM64 GB10/GX10 systems. Other Linux systems can contribute an existing supported local LLM server.

On a GB10/GX10, check the existing NVIDIA/container installation:

nvidia-smi
docker --version
nvidia-ctk --version
docker info

Resolve missing GPU/container prerequisites or Docker access before provisioning vLLM. Use the machine administrator's approved setup. Docker access grants broad host privileges; do not add users to the Docker group casually.

2. Install the commands

curl -fL --progress-bar --connect-timeout 20 https://github.com/mundusx/releases/releases/download/opengpu-prod/install.sh | bash

The script selects the architecture, downloads binaries/checksums, and prints progress. On detected GB10/GX10 it also prepares the pinned vLLM container; container layer downloads have their own progress display. This can take several minutes. Do not prefix the whole command with sudo.

If you will reuse an existing engine, use this command instead to skip automatic managed vLLM provisioning:

curl -fL --progress-bar --connect-timeout 20 https://github.com/mundusx/releases/releases/download/opengpu-prod/install.sh | bash -s -- --connection direct --cluster-url http://127.0.0.1:8000

Replace the URL with your engine's actual endpoint. The bootstrap prints the matching opengpu install command; it does not run the wizard automatically.

The default script installs first and leaves contributor choices to the wizard. If a command is not found, open a new terminal or run:

export PATH="$HOME/.local/bin:$PATH"

3. Prepare or discover the LLM runtime

For an existing vLLM, Ollama, LM Studio, llama.cpp, or supported OpenAI-compatible server, keep it running and inspect discovery:

opengpu cluster scan
opengpu install --connection direct --cluster-url http://127.0.0.1:8000

Choose the exact model in setup. For a noninteractive selection, provide its ID:

opengpu install --connection direct --cluster-url http://127.0.0.1:8000 --cluster-model YOUR_MODEL_ID

The URL above is an example, not a requirement. Use the actual listening address. A service listening on port 8000 is not automatically an LLM; ComfyUI or another service on that port must not be selected as one. Detection validates the model API. LM Studio must report the selected model as loaded. See Inference connections. MundusX does not provision Sparkrun recipes, and PAIR currently supports validation only.

For managed GB10/GX10 vLLM, run opengpu install --connection managed instead and choose an offered standalone model. For general Linux x64, do not assume the installer also installed a local LLM server; prepare a supported server before choosing that contribution path.

4. Configure, start, and verify

Finish the shared setup choices, then run:

opengpu --version
opengpu start --background
opengpu status
opengpu doctor
opengpu model list

This release reports opengpu 0.2.23. Complete onboarding if requested. Confirm readyForJobs: yes and the expected model/connection state. Inspect any reported failure reason before considering the machine connected and ready. Keep an external model server running while contributing it. Rerun setup if you change its model or endpoint.

5. Pause or disconnect

opengpu pause
opengpu resume
opengpu exit

After reboot, start the external engine and load its model first when using a direct connection, then run opengpu start --background. Managed setup reloads its saved active model. Use opengpu logs for logs and opengpu doctor for diagnostics. For optional image/video setup, continue to media setup.

Installation progress

The initial script download shows curl progress. The installer then prints its current step and download progress. Slow download, GPU checks, and Docker steps print an elapsed-time update every 10 seconds; elapsed time is not a percentage or an estimate of remaining time. A step failure prints its exit status. Do not launch a second installer while one is still running.