Run a GPU node

Nodes do two jobs. Serving nodes embed and serve the index; they earn 20% of every paid query, split by the chunks they serve. Training nodes run the LoRA fine-tune chosen by the round.

Today the index is embedded and served by one operator node. Training nodes register now and are assigned the first job when round 1 closes.

Requirements

• An NVIDIA GPU with CUDA and enough memory for the base model of the job: 12 GB for Qwen2.5 0.5B, 16 GB for 1B to 1.5B models, 20 GB or more for SmolLM2 1.7B.
• Python 3.10 or newer with torch, transformers, peft, datasets.
• Node.js 20 or newer for the worker.

Register and run

curl -O https://<this site>/sdk/nodal-node.mjs
curl -O https://<this site>/sdk/nodal-train.py
npm i viem
pip install torch transformers peft datasets accelerate

NODAL_KEY=0x… GPU="RTX 4090" VRAM=24 node nodal-node.mjs

The worker signs a registration with the key of your payout address, then sends a signed heartbeat every minute so the registry shows the node online. When a job is assigned, it pulls the dataset mix as JSONL and runs nodal-train.py (LoRA rank 16, alpha 32, loss on answers only).

You can also register from the network page with a browser wallet.