The SolanaLM vocabulary.
The terms you'll meet across the architecture, features, and quickstart, defined plainly.
- Gateway
- The FastAPI service that authenticates callers, enforces rate limits, sanitizes input, writes audit logs, exposes Prometheus metrics, and routes each request to a healthy node.
- Node registry
- The discovery and health-check layer that tracks node capacity and role, and fails over between nodes using circuit breakers.
- Inference node
- A node that serves LLM completions via a pluggable backend — local PyTorch/Transformers, llama.cpp, or a proxied provider.
- Training node
- A node that contributes GPU cycles to federated learning rounds and earns SOL from round participation.
- Federated learning (FL)
- Training a shared model across nodes that keep their data local, aggregating only model updates. SolanaLM implements FedAvg, FedProx, FedAdam, and SCAFFOLD.
- FedAvg
- Federated Averaging — the baseline aggregation strategy that averages client model weights each round.
- FedProx
- A federated strategy adding a proximal term to stabilize training under heterogeneous (non-IID) client data.
- SCAFFOLD
- A federated strategy that uses control variates to correct client drift, improving convergence on non-IID data.
- Differential privacy (DP)
- A privacy guarantee that adds calibrated noise so individual training records cannot be reconstructed from model updates.
- Onion routing
- Optional 3-hop routing for inference so no single node sees both the request source and the prompt content.
- Settlement
- Per-request SOL micropayment via a signed Solana transaction, with ~0.4s finality and ~$0.00025 fees.
- OpenAI compatibility
- The gateway implements OpenAI’s /v1/chat/completions and /v1/completions schema, so the OpenAI SDK works by changing only the base URL and key.