Module execution_plan 

Load-time execution / memory plans (llama.cpp graph-params analogue).

Selected once per model (and cached by batch geometry for decode vs prefill). Hot-path forward never re-derives architecture strings or fused-op availability.

Structs

ExecutionPlan

Per-model execution plan: everything the forward path needs that is constant across tokens.

FusedOpCaps

Backend fused-op availability discovered at load / first probe.

MemoryPlan

Memory layout chosen once from the architecture profile.

PlanGeometry

Cache key for decode/prefill plan reuse (batch geometry only).