Crate ghostlink_core 

Ghost-Link Core Library

A zero-config LAN fabric that turns spare local GPUs into a shared execution surface for large-model inference and training.

Features

• Zero-copy SPSC ring buffers for DMA-style hand-off
• Binary protocol with CRC32 checksums for frame integrity
• Thread-safe cluster state with metrics collection
• Greedy layer assignment with fault tolerance
• Experimental XDP scaffolding (currently unavailable in this build)
• Network health monitoring
• Load balancing and tensor distribution

Architecture

crates/
├── ghostlink-core/  # Shared runtime primitives
│   ├── ring.rs          # Zero-copy SPSC ring buffer
│   ├── protocol.rs      # Binary protocol with CRC32
│   ├── cluster.rs       # Thread-safe node state
│   ├── planning.rs      # Layer assignment + fault tolerance
│   ├── health.rs        # Network health monitoring
│   ├── load_balance.rs  # Tensor distribution
│   ├── kv_cache.rs      # KV cache primitive (no current caller — see module docs)
│   ├── xdp.rs           # Experimental XDP scaffolding (stubbed)
│   └── dashboard.rs     # Terminal UI with ratatui
└── ghost-link/          # CLI demo entrypoint

Re-exports

pub use accelerator::ExecutionBackend;

pub use autotune::AutoTuner;

pub use cluster::ClusterState;

pub use cluster::NodeMetrics;

pub use cluster::NodeStatus;

pub use discovery::broadcast_and_collect;

pub use discovery::UdpDiscoveryConfig;

pub use discovery::DEFAULT_DISCOVERY_PORT;

pub use host::detect_local_node_resources;

pub use host::detect_runtime_profile;

pub use host::detect_runtime_profile_full;

pub use host::detect_runtime_profile_with_mode;

pub use host::infer_compute_capability_from_name;

pub use host::is_apple_silicon;

pub use host::AccelerationMode;

pub use host::GpuBackend;

pub use host::ProbeMode;

pub use host::RuntimeProfile;

pub use kv_cache::KVCacheConfig;

pub use kv_cache::KVCacheEntry;

pub use kv_cache::LayerKvCache;

pub use planning::assign_layers_sequentially;

pub use planning::assign_layers_with_fault_tolerance_and_runtime;

pub use planning::assign_layers_with_runtime_profile;

pub use planning::chunk_assignments_for_workers;

pub use planning::select_quantization_mode;

pub use planning::LayerAssignment;

pub use planning::LayerSpec;

pub use planning::PlacementPlan;

pub use planning::PlanningTuning;

pub use planning::QuantizationMode;

pub use protocol::NodeResources;

pub use ring::RingConfig;

pub use ring::SpscRingBuffer;

pub use runtime::build_token_schedule;

pub use runtime::execute_pipeline;

pub use runtime::execute_pipeline_tcp_loopback;

pub use runtime::execute_pipeline_tcp_loopback_with_config;

pub use runtime::BridgeAddr;

pub use runtime::BridgeListener;

pub use runtime::BridgeStream;

pub use runtime::DeviceKind;

pub use runtime::ExecutionResult;

pub use runtime::PipelinePlan;

pub use runtime::StageExecutionStats;

pub use runtime::StagePlacement;

pub use runtime::TcpTransportConfig;

pub use runtime::TokenStep;

pub use runtime::TransportKind;

pub use system_profile::CpuFeatures;

pub use system_profile::CpuInfo;

pub use system_profile::GpuInfo;

pub use system_profile::MemoryInfo;

pub use system_profile::NetworkInfo;

pub use system_profile::NpuInfo;

pub use system_profile::SystemProfile;

pub use watcher::ProfileChange;

pub use watcher::SystemProfileWatcher;

Modules

accelerator

Runtime-selected execution backend helpers.

autotune

Unified auto-tuning for Ghost-Link subsystems.

cluster

Thread-Safe Cluster State with Metrics Collection

dashboard

Terminal Dashboard with ratatui Integration for Ghost-Link Cluster

discovery

UDP discovery fallback for mixed LAN environments.

health

Network Health Monitoring for Ghost-Link Cluster

host

Host resource detection and runtime auto-tuning.

kv_cache

KV cache primitives for a single sequence within a single transformer layer.

load_balance

Tensor Distribution and Load Balancing for Ghost-Link Cluster

mdns

mDNS-based peer discovery — an alternative path to the UDP broadcast fallback in discovery.rs for networks where broadcast traffic is filtered but multicast is not (some managed VLANs, cloud VPCs). Runs alongside UDP discovery rather than replacing it: callers typically merge results from both.

models

Model Registry and Asset Management

planning

Greedy Layer Assignment with Fault Tolerance and Adaptive Quantization

protocol

Binary Protocol with CRC32 Checksums for Ghost-Link Discovery

ring

Zero-Copy SPSC Ring Buffer with AF_XDP Integration

runtime

Distributed inference runtime scaffolding.

system_profile

Unified system profile: single source of truth for host capabilities.

watcher

Dynamic system-profile watcher for hot-plug hardware detection.