Struct SubBackend 

pub struct SubBackend { /* private fields */ }

A backend view that restricts communication to a subset of global ranks.

SubBackend wraps a parent Backend and a list of member global ranks to form a logical sub-process-group. It translates every rank(), world_size(), send(), and recv() call into the equivalent parent- backend operation using the subgroup→global rank mapping.

Because SubBackend implements the Backend trait, the existing allreduce, all_gather, and reduce_scatter collective functions work on a subgroup without any changes: they only see the SubBackend through the trait, so “rank 0” in the collective means “the first member of the subgroup” and world_size means the subgroup size.

Used by FSDP’s HybridShard strategy to form an intra-node (sharding) subgroup and an inter-node (replication) subgroup.

CL-327.

Implementations

impl SubBackend

pub fn new( parent: Arc<dyn Backend>, members: Vec<usize>, ) -> FerrotorchResult<Self>

Create a subgroup view from a parent backend and a list of member global ranks.

The caller’s rank (read from parent.rank()) must be in members. members is sorted and deduplicated before being stored.

Errors

• DistributedError::InvalidRank if the parent’s rank is not in members, or if any member is ≥ parent world_size.
• DistributedError::InvalidWorldSize if members is empty.

pub fn members(&self) -> &[usize]

Return the global ranks that make up this subgroup, in ascending order. The index of this rank’s entry is its local rank.

pub fn to_global(&self, local: usize) -> usize

Map a local (subgroup-relative) rank to its global rank.

Panics

Panics if local is out of bounds.

pub fn to_local(&self, global: usize) -> Option<usize>

Map a global rank to its local subgroup rank, or None if the global rank is not a member of this subgroup.

pub fn parent(&self) -> &Arc<dyn Backend>

The parent backend this subgroup was derived from.

Trait Implementations

impl Backend for SubBackend

fn rank(&self) -> usize

This rank’s index in the world (0-based).

fn world_size(&self) -> usize

Total number of ranks in the process group.

fn send(&self, data: &[u8], dst_rank: usize) -> FerrotorchResult<()>

Send data to dst_rank.

fn recv(&self, dst: &mut [u8], src_rank: usize) -> FerrotorchResult<()>

Receive into dst from src_rank. The caller must allocate dst with the correct length before calling.

fn recv_timeout( &self, dst: &mut [u8], src_rank: usize, timeout: Duration, ) -> FerrotorchResult<()>

Receive into dst from src_rank with a timeout.

fn barrier(&self) -> FerrotorchResult<()>

Block until every rank has reached this barrier.