Trait AddressableExpertRouteProvider 

pub trait AddressableExpertRouteProvider<T> {
    type Error;

    // Required method
    fn execute_addressable_routes(
        &mut self,
        request: AddressableExpertRouteRequest<'_, T>,
    ) -> Result<T, Self::Error>;

    // Provided method
    fn execute_addressable_routes_tensor_parallel(
        &mut self,
        request: AddressableExpertRouteRequest<'_, T>,
    ) -> Result<RoutedExpertTensorParallelOutput<T>, Self::Error> { ... }
}

Local addressable grouped execution used by expert exchange.

The provider must consume every submitted row exactly once, select its corresponding owner-local expert, and apply its route coefficient exactly once. Acquired bank resources remain provider-owned until the returned tensor is natively complete.

Required Associated Types

type Error

Acquisition or grouped execution failure.

Required Methods

fn execute_addressable_routes( &mut self, request: AddressableExpertRouteRequest<'_, T>, ) -> Result<T, Self::Error>

Executes one owner-local grouped batch.

Provided Methods

fn execute_addressable_routes_tensor_parallel( &mut self, request: AddressableExpertRouteRequest<'_, T>, ) -> Result<RoutedExpertTensorParallelOutput<T>, Self::Error>

Executes one owner-local grouped batch while retaining tensor-parallel reduction structure.

Providers without rank-local TP work inherit complete-output behavior. A TP provider overrides this method and returns its reducible activation contribution plus the optional selection-weighted post-reduction bias. The exchange protocol returns both values to their source-token order; it must not add the bias before the caller’s tensor all-sum.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors