avenia-worker

Run a GPU provider for the Avenia network — a decentralized LoRA fine-tuning marketplace. Your box joins the hosted coordinator over WebSocket, receives training jobs, fine-tunes real LoRA adapters with HuggingFace + peft, uploads the result, and earns for the compute.

No GPU? The worker still runs against a built-in deterministic mock trainer (--runtime mock, the default), so you can join, list models, and exercise the whole flow with zero hardware.

Install on a fresh GPU box

On a rented GPU box (e.g. a RunPod PyTorch pod) you need Node 20+, then the worker, then the Python trainer dependencies.

# 1. Node 20+ (skip if the box already has it)
curl -fsSL https://deb.nodesource.com/setup_20.x | bash - && apt-get install -y nodejs

# 2. The worker CLI
npm install -g avenia-worker

# 3. The real trainer's Python deps (only needed for --runtime python)
pip install torch transformers peft datasets accelerate

# 4. Join the network and train real jobs
avenia-worker start \
  --runtime python \
  --models aven-qwen-0.5b \
  --account 0xYourWallet \
  --url wss://avenia.fly.dev/worker

--account is the wallet that receives your provider earnings.

Commands

avenia-worker start [--runtime mock|python] [--models a,b] [--account 0x..] [--region r] [--url wss://...] [--label bootstrap|marketplace] [--vram gb]
avenia-worker list       [--runtime mock|python] [--vram gb]    # models this box can serve
avenia-worker benchmark  [--runtime mock|python] [--vram gb]    # measure samples/sec per model

Environment:

• AVENIA_RUNTIME=mock|python overrides --runtime (default: mock).

Runtimes

• mock (default) — a deterministic synthetic trainer. No GPU, no Python.
• python — spawns the bundled trainer/train_lora.py / trainer/infer_lora.py sidecars (HuggingFace + peft). Requires a GPU/MPS and the deps from step 3 above (or pip install -r $(npm root -g)/avenia-worker/trainer/requirements.txt).

License

MIT