Home

Welcome to the Ghostlink Wiki

Welcome to the official developer and user documentation for Ghostlink (currently v1.16.1). Ghostlink is a high-performance LAN fabric designed to turn spare local GPUs, CPUs, and NPUs into a unified, low-latency execution surface for large-model inference.

Use the navigation below or the sidebar to explore technical guides, deployment strategies, and architecture deep-dives.

📖 Documentation Directory

🚀 Getting Started

Quickstart Guide: The fastest path to spin up your first cluster and launch Ghostlink Studio.
Installation & Requirements: Hardware prerequisites (CPU, GPU, NPU, DirectML, Metal) and environment setup.
CLI Reference: Breakdown of core commands (serve, probe, plan, flow, doctor).

🏗️ Architecture & Core Primitives

Architecture Overview: Component hierarchy, control-plane vs data-plane, and the reactive fabric.
Network & Discovery Layer: Zero-conf mDNS discovery, wireguard-in-user-space topology, and heartbeat mechanics.
Layer Placement & Execution Planning: Dynamic heuristics for tensor splitting, VRAM offloading, and priority queuing.

🌐 Deployment & Operations

Multi-Node Deployment Guide: Production staging, network autotuning, and routing across complex LAN fabrics.
Security Model: Transport hardening, PQC/hybrid TLS, JWT auth, and audit logging.
Troubleshooting & Diagnostics: Using cargo run -p ghost-link -- doctor --strict to debug environment drift.

🛠️ Service Port Architecture

Ghostlink operates across a three-tier local service topology:

Port	Component	Role
8000	Control-Plane Gateway (Go)	GUI-facing entry point (CORS, rate limiting, logging, streaming proxy)
8003	Ghostlink API Server	Internal engine (Chat, models, settings, system profile)
8080	Native Inference Engine	`llama-server` or Ollama backend execution

Note: The React GUI (http://127.0.0.1:5173) must always connect through the Control-Plane Gateway on port 8000.

📊 Expected Performance Baselines

Transport-layer throughput measured on an Intel i7-14700K (Linux/WSL2) compiled with target-cpu=native:

Transport Mode	Tokens / Batch	Throughput	Latency	Target Context
In-Process (Spin-Wait)	1024 / 128	900 K tok/s	1.14 ms	Co-located processes / Same host
TCP Loopback	1024 / 128	340 K tok/s	3.01 ms	Local containers & IPC bridges
In-Process (Spin-Wait)	256 / 32	639 K tok/s	0.40 ms	Low-latency micro-batches
TCP Loopback	256 / 32	236 K tok/s	1.08 ms	Local inter-process stream

⚠️ Note: Microbenchmarks move 64-byte synthetic payloads. For realistic FP16/BF16 LLM activation layer transfers across physical LAN nodes, see docs/BENCHMARKS.md in the repository.

📟 CLI Command Cheatsheet

Standard Full-Stack Launch

# Windows
.\launch.bat

# Linux / macOS
./launch.sh

Navigation Menu