memory
Memory planning API.
The public interfaces describe tensor allocation constraints, arena placement and scratch diagnostics. The default hill_climb planner, the greedy planner and the experimental greedy_by_size planner consume the same resolved model and constraints. A plan defines module arena requirements; it does not reserve memory for the application or prove an inference-time improvement.
Use the memory guides for placement, planner selection and residency reports. The public re-exports below link to their defining contracts; implementation modules not present in the manifest are not additional API pages.
The configuration types (MemoryConstraint, MemoryPlannerType and
TensorAttributes) are promised. The planner’s output types (MemoryPlan,
TensorAllocation, TensorBinding, TensorLifetime, ArenaUsage and
ArenaRole) are provisional and can change in a minor release; the residency
report is the supported reader of a plan.
Re-exports
| Name | Source |
|---|---|
| ArenaRole | helia_aot.memory.defines.ArenaRole |
| ArenaUsage | helia_aot.memory.defines.ArenaUsage |
| MemoryConstraint | helia_aot.memory.defines.MemoryConstraint |
| MemoryPlan | helia_aot.memory.defines.MemoryPlan |
| MemoryPlannerType | helia_aot.memory.defines.MemoryPlannerType |
| TensorAllocation | helia_aot.memory.defines.TensorAllocation |
| TensorAttributes | helia_aot.memory.defines.TensorAttributes |
| TensorBinding | helia_aot.memory.defines.TensorBinding |
| TensorLifetime | helia_aot.memory.defines.TensorLifetime |