Struct LayeredPartitionDriver 

pub struct LayeredPartitionDriver { /* private fields */ }

Validated execution metadata for one rank-local layered partition.

This driver is the single owner of partition input/output checks, canonical storage and state ranges, execution-group setup/completion, and final output projection. Concrete backends retain only state storage and unit residency.

Implementations

impl LayeredPartitionDriver

pub fn new<G, A>( partition: &ArchitecturePartition<G, A>, group_index: usize, storage_range: Range<usize>, ) -> Result<Self, LayeredPartitionError>

Validates a canonical partition against its concrete unit storage.

pub fn range(&self) -> Range<usize>

Returns the canonical group-local global unit range.

pub const fn group_index(&self) -> usize

Returns the canonical architecture execution-group slot.

pub const fn state_layout(&self) -> &StateLayout

Returns the architecture-global state layout for this partition.

pub fn input<'a, T, A>( &self, input: LayeredPartitionInput<'a, T, A>, ) -> Result<LayeredPartitionInput<'a, T, A>, LayeredPartitionError>

Validates input form against architecture boundary ownership.

pub fn exchange_boundary<B>( &self, value: B::Tensor, group: &B::Group, executor: &B::Executor, ) -> Result<B::Tensor, B::CollectiveError>

where B: CollectiveBackend,

Transports this partition’s realized boundary through the selected opaque collective group.

Keeping the operation on the validated driver makes boundary movement part of partition execution rather than an unrelated backend call.

pub fn begin<'a, B, S, M>( &self, architecture: &mut M, input: LayeredPartitionInput<'a, B::Tensor, <M::Boundary as ArchitectureBoundary>::Boundary<B::Tensor>>, mask: Option<&B::Tensor>, state: &mut S, parallel: Option<&B::ParallelContext>, context: &<B::Tensor as Tensor>::Context, ) -> Result<LayeredForwardState<B::Tensor, M::ForwardContext>, M::Error>

where B: NeuralBackend, S: RuntimeState<B>, M: PartitionedLayeredArchitecture<B, S>,

Prepares the partition and starts its canonical execution group.

pub fn finish<B, S, M>( &self, architecture: &mut M, hidden: &B::Tensor, state: &mut S, forward: &mut M::ForwardContext, parallel: Option<&B::ParallelContext>, context: &<B::Tensor as Tensor>::Context, ) -> Result<LayeredPartitionOutput<B::Tensor, <M::Boundary as ArchitectureBoundary>::Boundary<B::Tensor>>, M::Error>

where B: NeuralBackend, S: RuntimeState<B>, M: PartitionedLayeredArchitecture<B, S>,

Completes the canonical group and applies output projection only on its owner.

Trait Implementations

impl Clone for LayeredPartitionDriver

fn clone(&self) -> LayeredPartitionDriver

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for LayeredPartitionDriver

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.