Crate neuronbox_runtime 

NeuronBox runtime library: GPU detection, daemon protocol, session tracking.

Re-exports

pub use gpu::detect_gpus;

pub use gpu::primary_gpu_vram_mb;

pub use gpu::soft_vram_check;

pub use gpu::GpuDevice;

pub use gpu::GpuError;

pub use host::HostProbe;

pub use host::HostSnapshot;

pub use host::TrainingBackend;

pub use protocol::ActiveModelInfo;

pub use protocol::DaemonRequest;

pub use protocol::DaemonResponse;

pub use protocol::SessionInfo;

Modules

gpu

GPU detection and soft VRAM control — delegates probing to crate::host.

gpu_manager

In-memory session registry (soft VRAM accounting per registered PID).

host

Single host probe layer: GPU, platform, likely PyTorch backend.

model_loader

Tracks logical active model for swap / display (weights are loaded by Python / user code, not in Rust).

protocol

server

Unix socket server (newline-delimited JSON).

vram_watch

Soft monitoring: if VRAM usage reported by nvidia-smi for a PID exceeds ~115% of the registered estimate, send SIGKILL (Linux only).