Power Node

Distributed GPU compute node for the Gelotto AI image generation network.

Power Node lets you contribute GPU compute power to generate AI images and earn rewards. It connects to the Gelotto network, claims generation jobs, and uses your NVIDIA GPU to create images.

Requirements

• NVIDIA GPU with 8GB+ VRAM (RTX 3060, RTX 4060, or newer)
• Linux with NVIDIA drivers installed
• Python 3.10+
• CUDA toolkit (for GGUF mode, to build stable-diffusion.cpp)
• Disk space:
  • GGUF mode: ~15GB (base) + ~500MB (face-swap)
  • PyTorch mode: ~40GB (base) + ~29GB (video, optional) + ~500MB (face-swap)

Supported GPUs

GPU	VRAM	Mode	Video	Disk Space	Performance
RTX 3060	12GB	GGUF	No	~15GB	Good
RTX 4060	8GB	GGUF	No	~10GB	Good
RTX 4070/Ti	12GB	GGUF	No	~15GB	Better
RTX 4080	16GB	GGUF	No	~15GB	Better
RTX 4090	24GB	GGUF	No	~15GB	Best
RTX 5070 Ti	16GB	PyTorch	Yes	~70GB	Best
RTX 5080	16GB	PyTorch	Yes	~70GB	Best
RTX 5090	32GB	PyTorch	Yes	~70GB	Best

Notes:

• RTX 50-series (Blackwell) GPUs require PyTorch mode due to CUDA compute capability 12.0
• Video generation requires PyTorch mode with 12GB+ VRAM (automatically detected)

Quick Start

1. Register Your Worker

Visit https://gelotto.io/workers and register with a unique hostname.

You'll receive:

• Worker ID - Your unique identifier
• API Key - Your authentication token (save this securely!)

2. Install Power Node

Option A: One-line installer (recommended)

curl -sSL https://raw.githubusercontent.com/Gelotto/power-node/main/install.sh | bash

Option B: Clone and install

git clone https://github.com/Gelotto/power-node.git
cd power-node
./install.sh

The installer will:

1. Detect your GPU and VRAM
2. Download the appropriate model (~9GB for GGUF, ~31GB for PyTorch)
3. Auto-download optional models based on GPU capability:
   • Video generation (Wan2.1, ~29GB) - PyTorch mode with 12GB+ VRAM
   • Face-swap (~500MB) - Any mode with 6GB+ VRAM
4. Set up Python environment and dependencies
5. Create configuration files

Installation time: 10-30 minutes depending on internet speed (longer if video model is downloaded)

Installation Options

Skip optional models with these flags:

# Skip video model only
curl -sSL ... | bash -s -- --no-video

# Skip face-swap only
curl -sSL ... | bash -s -- --no-faceswap

# Skip all optional models (minimal install)
curl -sSL ... | bash -s -- --minimal

The installer will warn if disk space is below 50GB when downloading the video model.

3. Configure

Edit ~/.power-node/config/config.yaml and add your credentials:

api:
  key: "YOUR_API_KEY_HERE"

worker:
  id: "YOUR_WORKER_ID_HERE"

4. Start

~/.power-node/start.sh

You should see:

Starting worker...
Using credentials - Worker ID: abc123...
API Key: wk_xxxxxxxxxx...
Starting Python inference service...
Worker fully initialized!
Starting job loop (polling every 5s)...

5. Run as System Service (Optional)

To start automatically on boot:

sudo cp ~/.power-node/power-node.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable power-node
sudo systemctl start power-node

Check status:

sudo systemctl status power-node
sudo journalctl -u power-node -f

How It Works

1. Register - Create an account at gelotto.io/workers
2. Connect - Power Node connects to the Gelotto API
3. Claim Jobs - The node polls for available generation jobs
4. Generate - Your GPU generates images using the Z-Image model
5. Upload - Results are uploaded back to the network
6. Earn - Receive rewards for completed jobs

Configuration Reference

api:
  url: https://api.gelotto.io      # API endpoint
  key: ""                            # Your API key

model:
  service_mode: gguf                 # gguf or pytorch (auto-detected)
  vram_gb: 8                         # Your GPU VRAM

worker:
  id: ""                             # Your worker ID
  hostname: ""                       # Auto-detected if empty
  gpu_info: ""                       # Auto-detected if empty
  poll_interval: 5s                  # Job polling interval
  heartbeat_interval: 30s            # Heartbeat interval

python:
  executable: ~/.power-node/venv/bin/python3
  script_path: ~/.power-node/scripts/inference.py
  script_args: []                    # Model paths (auto-configured)
  env:
    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"

Troubleshooting

GPU not detected

Ensure NVIDIA drivers are installed:

nvidia-smi

Out of memory errors

The node automatically configures for your GPU's VRAM. For 8GB GPUs:

• Uses quantized Q4_0 models
• Enables VAE tiling
• Offloads components to CPU

If you still get OOM errors, try reducing resolution or closing other GPU applications.

Connection errors

1. Check your API key and worker ID are correct in config.yaml
2. Verify network connectivity: curl https://api.gelotto.io/health
3. Check if the API is reachable from your network

Model download failed

If model download was interrupted:

rm -rf ~/.power-node/models
~/.power-node/install.sh

Python errors

Check Python environment:

source ~/.power-node/venv/bin/activate
python3 -c "import torch; print(torch.cuda.is_available())"

Building from Source

Requires Go 1.21+:

git clone https://github.com/Gelotto/power-node.git
cd power-node
go build -o bin/power-node ./cmd/power-node

Environment Variables

Variable	Description	Default
POWER_NODE_DIR	Installation directory	~/.power-node
API_URL	Backend API URL	https://api.gelotto.io

License

MIT License - see LICENSE for details.

Links

• Gelotto Network
• PicShapes AI Generator
• Worker Registration
• Discord