Module offload 

CPU/disk offload with a pinned hot-layer set.

This module provides the infrastructure for offloading model weights to disk and loading them on-demand with an LRU eviction policy. A pinned hot-set keeps embeddings, the output head, and the last N attention layers always resident in RAM.

Overview

• OffloadPolicy — declarative configuration: none, budget, or pinned hot-set.
• LayerPager — LRU weight pager with eviction, pinned tensors, and on-demand loads from a PagerSource.
• MemoryPressureProbe — lightweight OS-level pressure monitor (Linux / macOS).

Re-exports

pub use pager::FilePagerSource;

pub use pager::LayerPager;

pub use pager::PagerSource;

pub use pager::ResidentTensor;

pub use pager::TensorEntry;

pub use pager::TensorId;

pub use policy::OffloadPolicy;

pub use pressure::MemoryPressureProbe;

Modules

pager

LRU weight pager — the core of the CPU/disk offload system.

policy

Offload policy configuration.

pressure

Memory pressure probe — lightweight OS-level RAM usage monitor.