Backend Availability

Build-dependent availability for backend capsules.

KernelPolicy says which capsules selection may use. Availability says which devices and providers a build declares. CPU-only and LibTorch CUDA builds share one semantic registry while exposing different capsule subsets to the planner. An availability declaration is not runtime discovery; executable paths still probe the linked runtime before launching work.

structure NN.Backend.Availability : Type

Devices and providers declared available to backend planning.

• devices : Array Device
• providers : Array Provider

def NN.Backend.instReprAvailability.repr : Availability → Nat → Std.Format

@[instance_reducible]

instance NN.Backend.instReprAvailability : Repr Availability

def NN.Backend.Availability.admitsCapsule (a : Availability) (c : KernelCapsule) : Bool

Whether a capsule can even be considered on this build.

def NN.Backend.Availability.filterCapsules (a : Availability) (capsules : Array KernelCapsule) : Array KernelCapsule

Keep only capsules available on this build.

def NN.Backend.Availability.cpu : Availability

CPU/reference-only availability.

def NN.Backend.Availability.cuda : Availability

CUDA availability with LibTorch primitives and TorchLean compositions of those primitives.