Module expert

Expand description

Runtime ownership boundary for routed expert acquisition and residency.

Structs§

AddressableBankBindingPlan: Canonical binding and transformation plan for one addressable member.
AddressableBankMember: Exact generic storage member projected from an architecture-owned bank catalog.
AddressableBankMemberPlacement: Architecture-selected ownership retained with an addressable member.
AddressableBankParameter: One exact selected parameter in an independently addressable bank member.
AddressableBindingTransform: Backend-neutral transform retained for one addressable binding.
AddressableExpertRouteRequest: One owner-local grouped batch submitted after expert exchange.
ObservedExpertProvider: Decorates a routed provider with normalized routing observation.
ParameterBankAcquisition: Exact generic request for an independently addressable bank acquisition.
ResidentExpertProvider: Provider for a fully resident expert bank.
RoutedExpertRequest: One architecture route batch submitted to a runtime expert provider.
RoutedObservationPoint: Stable routing metadata supplied by an architecture composition at one canonical unit boundary.

AddressableBankDistribution: Neutral placement class for one independently addressable member.
AddressableBankMemberError: Invalid generic addressable-bank member projection.
ExpertRouteCombination: Architecture-selected combination for one expert-exchange batch.
ObservedExpertProviderError: Failure from either canonical expert execution or its observation hook.
RoutedExpertTensorParallelOutput: Provider result that distinguishes complete outputs from rank-local TP work.

AddressableExpertRouteProvider: Local addressable grouped execution used by expert exchange.
AddressableGatedProductBank: Mechanism-only lookup of one grouped operator in an addressable parameter bank.
AddressableGroupedBank: Generic addressable storage and grouped-operator construction mechanisms.
ExpertRouteExchange: Opaque variable-count transport used by architecture-owned expert routing.
ExpertRouteTensorMovement: Backend-neutral tensor movement needed by an expert-exchange protocol.
IndexedMovement: Generic indexed tensor movement required by bounded grouped execution.
RoutedExpertProvider: Runtime boundary for resident or independently cached routed experts.
TensorParallelRoutedExpertProvider: Additive provider mechanism for tensor-parallel grouped partials.

combine_routed_expert_tensor_parallel: Combines routed/shared provider outputs while requiring one coherent TP mode.
combine_tensor_parallel_expert_outputs: Combines two rank-local expert partials without introducing another collective.
plan_addressable_bank_bindings: Validates exact addressable tasks and derives their singular canonical binding plans.
reduce_routed_expert_tensor_parallel: Completes a provider TP result while preserving provider-owned collectives.
reduce_tensor_parallel_expert_output: Completes one rank-local expert output with one all-sum and one post-bias add.
select_routes_with_observer: Selects with a direct architecture observation hook, preserving shared-expert work outside routed control. No unmodified model forward pass is performed.
select_routes_with_provider: Selects through an architecture-owned selector before an observed provider’s expert call. Ordinary providers take the original lightweight select path.
selected_addressable_parameter_bytes: Computes executable storage bytes for one admitted member-local task output.