How psychefoundation/psyche works
Project type and purpose
• Psyche is a monorepo for a distributed, transformer‐based ML training framework. It provides both (a) a “centralized” reference architecture (Rust clients, server, local testnet) and (b) a “decentralized” variant built on Solana Anchor programs (authorizer, coordinator, mining‐pool, treasurer) plus off‐chain tooling, Python bindings, a web frontend/backend, telemetry, Docker/Nix infrastructure, documentation, and shared libraries.
File structure
• Top-level “architectures” folder splits into:
· centralized: client, server, shared protocol, local testnet, testing harness
· decentralized: on-chain programs (authorizer, coordinator, mining-pool, treasurer), client CLI, tooling, chaos/testing
• “shared” folder: core libraries for client, coordinator, data‐provider, modeling, network (P2P), metrics, eval, tui, watcher
• “python” folder for Python extension (PyO3 sidecar), dtensor helpers, high-level APIs
• “website” folder: Node.js backend (TypeScript + MikroORM), React/TypeScript frontend, shared TypeScript types, WebAssembly (Rust→wasm)
• “telemetry”: Prometheus, Grafana, Loki, OpenTelemetry collector configs
• “docker”, “nix”, “scripts”, “tools”: build & deployment automation
README insights
• Core goal: enable trustless, collaborative training over the Internet via off‐chain compute nodes and on‐chain coordination & reward distribution.