Project

go-exotic

experimental

Go experiment for deterministic model sharding, route planning and local layer-range execution.

Overview

go-exotic explores distributed inference planning in Go, inspired by exo. It uses go-pherence as the local runtime and separates deterministic placement from actual shard execution.

The current project is an orchestration and validation experiment, not a networked distributed-inference service.

How it works

The CLI and local web server expose peer capabilities, memory-weighted layer placement and route previews. An in-process simulator checks orchestration before any execution path is enabled.

The shard endpoint is disabled by default. Supplying a local model installs a go-pherence layer executor for bounded local development; it does not enable peer discovery or LAN generation.

Features
Deterministic placement

Assigns model layers to advertised peer memory with stable, inspectable plans.

Route previews

Shows the proposed layer and peer path without executing shards or modifying local models.

Local runtime adapter

Uses go-pherence for model metadata, tokenisation, smoke generation and layer-range execution.

In-process simulation

Exercises orchestration and failure paths without requiring several machines.

Gated execution

POST /shards/execute returns 503 shard execution disabled until a local executor is explicitly configured.

Dashboard and API

A Bun-built Preact/D3 dashboard presents status, capabilities, placement, routes and bounded local model inventory.

Architecture
CLI / dashboard plan · peers · routes · run Go service local HTTP API Placement planner memory-weighted layers Simulator routes · orchestration checks Local executor go-pherence · opt-in Preview / result JSON · D3 dashboard Deterministic shard planning with gated local execution