Backend Registry

Registry of backend capsules known to TorchLean's planner.

Capsules are contributed by operation or provider modules and flattened into one planner catalog. Model architectures never appear here: they lower to backend operations, and the planner chooses a capsule for each operation. Build availability filters the maintained LibTorch CUDA capsules from CPU-only profiles. Caller-supplied modules still pass the same availability and assurance checks.

structure NN.Backend.Registry.CapsuleModule : Type

A named, independently maintained contribution to the backend catalog.

• name : String
• capsules : Array KernelCapsule

def NN.Backend.Registry.instReprCapsuleModule.repr : CapsuleModule → Nat → Std.Format

@[instance_reducible]

instance NN.Backend.Registry.instReprCapsuleModule : Repr CapsuleModule

def NN.Backend.Registry.flatten (modules : Array CapsuleModule) : Array KernelCapsule

Flatten capsule modules while preserving module and local preference order.

def NN.Backend.Registry.firstDuplicateModuleName? (modules : Array CapsuleModule) : Option String

First repeated module name, if the registry contains two contributions with the same name.

def NN.Backend.Registry.firstDuplicateCapsuleIdentity? (capsules : Array KernelCapsule) : Option String

First repeated capsule identity after flattening, if one exists.

def NN.Backend.Registry.validateModules (modules : Array CapsuleModule) : Except String Unit

Validate the identities that make registry ordering meaningful.

Different providers may implement the same operation. What is rejected is registering the same named module twice or repeating the same capsule name, operation, provider, and device tuple.

def NN.Backend.Registry.maintainedModules : Array CapsuleModule

Maintained operation/provider modules. A new architecture does not modify this list; only a new primitive implementation or provider does.