skainet-lang-core/sk.ainet.lang.tensor.storage/MemoryPlanner

MemoryPlanner

class MemoryPlanner(availableDevices: Set<DeviceKind> = setOf(DeviceKind.CPU))(source)

Resolves Placement intent into concrete buffer allocation decisions.

The planner inspects available backends and decides:

Where a tensor should actually live (device + memory domain)
Whether a fallback is needed (e.g. GPU not available → CPU)
Whether immutable weights should be file-backed vs heap-copied

Currently only the CPU backend is wired in, so the planner always resolves to CPU/HOST_HEAP or CPU/MMAP_FILE. GPU/NPU resolution will be added when those backends ship.

Constructors

MemoryPlanner

constructor(availableDevices: Set<DeviceKind> = setOf(DeviceKind.CPU))

Functions

resolve

fun resolve(requested: Placement): ResolvedPlacement

Resolve a placement intent to an actual placement that can be satisfied.

suggestActivationPlacement

fun suggestActivationPlacement(): Placement

Suggest placement for a mutable activation/intermediate tensor.

suggestWeightPlacement

fun suggestWeightPlacement(isFileBacked: Boolean): Placement

Suggest the best placement for a weight tensor. File-backed if persistent, heap if transient.