Orvix Node Software

A Python agent that runs on a GPU provider's machine. It connects to the Orvix Orchestrator over WebSocket, registers its GPU, receives inference jobs, runs them, and returns results — earning USDC for the provider.

Inference is mocked by default, so the whole pipeline runs on a machine with no GPU at all. Set backend: "vllm" to serve real traffic — that path is live, and image generation runs alongside it on the same card.

Hardware requirements

• For real inference: NVIDIA GPU, CUDA 11+, 8 GB+ VRAM (Linux).
• For development: anything — use ORVIX_NODE_STUB_GPU=true and the mock backend.

Installation

One-line (Linux providers):

curl -sSL https://raw.githubusercontent.com/OrvixCompute/orvix/main/orvix-node/install.sh | bash
orvix-node join            # paste the credentials from the dashboard
orvix-node start

Manual (development, any OS):

cd orvix-node
python -m venv .venv
# Windows: .venv\Scripts\Activate.ps1   |   Unix: source .venv/bin/activate
pip install -e .            # core only (mock backend)
# pip install -e .[nvml]    # + real GPU detection (no vLLM)
# pip install -e .[gpu]     # + vLLM for real inference (Linux/CUDA)

Verify:

orvix-node --version

Configuration

Create the config file:

orvix-node config init        # writes ~/.orvix/config.yaml
orvix-node config show        # prints resolved config (secrets masked)

Precedence: CLI flags > env vars (ORVIX_NODE_*) > config file > defaults. Required fields: provider_id, node_secret (get them from POST /v1/provider/register on the orchestrator).

Running

# Development without a GPU (mock everything):
ORVIX_NODE_STUB_GPU=true orvix-node start

# Check the GPU detector:
ORVIX_NODE_STUB_GPU=true orvix-node gpu
ORVIX_NODE_STUB_GPU=true orvix-node gpu --watch

# Run inference locally without the orchestrator:
orvix-node test-inference --prompt "Hello, world"
orvix-node test-inference --prompt "Stream this" --stream

# Live status (queries the local health endpoint):
orvix-node status

# Tail logs:
orvix-node logs --tail 100 --follow

The node exposes a local health server (default :9000):

• GET /health → status, uptime, current jobs, GPU health, orchestrator connection
• GET /metrics → counters + live GPU metrics

Running as a systemd service

The installer can set this up, or do it manually:

# /etc/systemd/system/orvix-node.service
[Service]
ExecStart=%h/.local/bin/orvix-node start
Restart=always

sudo systemctl enable --now orvix-node
systemctl status orvix-node

Connection flow

Node                                  Orchestrator
 │ ── WS connect /v1/node/connect ───────▶ │
 │ ── RegisterMessage ───────────────────▶ │  validate provider + secret
 │ ◀── RegisterAck(accepted, node_id) ──── │
 │                                          │
 │ ── Heartbeat (every 15s) ─────────────▶ │  status, current_jobs, GPU metrics
 │ ◀── JobMessage ──────────────────────── │  dispatched inference request
 │ ── JobResult / JobChunk(stream) ──────▶ │  result correlated to the job
 │ ◀── Ping / Shutdown ─────────────────── │

On disconnect the node reconnects with exponential backoff (1→2→4…→60s). A rejected registration (accepted=false) is not retried.

Releasing

Publishing is automated and tokenless — PyPI trusts this repository through OIDC, so there is no API token to leak.

# 1. bump orvix_node/version.py (the single source; pyproject reads it)
# 2. merge that
git tag node-v0.2.1
git push origin node-v0.2.1

The node- prefix matters: this repository also tags its own releases as v0.2.0, and the package has a separate version line. The prefix says which artefact moved.

The workflow refuses to publish if the tag disagrees with version.py. PyPI never lets a version be re-uploaded, so a mismatch is worth failing on rather than discovering afterwards.

Providers install the last published version. A fix merged to main does not reach them until a release is cut — set ORVIX_NODE_REF to install from a git ref if you need one before then.

Architecture

Local integration with the orchestrator

1. Run the orchestrator on :8000.
2. Point the node at it: ORVIX_NODE_ORCHESTRATOR_URL=ws://localhost:8000.
3. Start the node (ORVIX_NODE_STUB_GPU=true orvix-node start).
4. Send a request via the OpenAI client to the orchestrator — it routes to the node.

Testing

pip install -e .[dev]
pytest -q

# Standalone client smoke test against an in-process mock server:
ORVIX_NODE_STUB_GPU=true python test_connection.py

Troubleshooting

• No GPU detected — pip install --upgrade "orvix-node[nvml]", or set ORVIX_NODE_STUB_GPU=true for development.
• Refusing insecure ws:// — only ws://localhost is allowed without TLS; use wss:// for remote orchestrators.
• Auth failed (exit 2) — check provider_id / node_secret against the orchestrator's /v1/provider/register.

Status

Job routing, provider earnings and USDC withdrawals are live, and vLLM inference and image generation both run in production. Providers are paid a share of each job they serve.

Staking is disabled during alpha, so the provider stake requirement is not enforced yet. Expect breaking changes.