Struct TcpTransport 

pub struct TcpTransport { /* private fields */ }

TCP-based transport for multi-node gradient exchange.

One node acts as coordinator (rank 0, TCP server) while all other nodes act as workers (TCP clients). The coordinator calls TcpTransport::coordinator which binds to a socket and blocks until world_size - 1 workers have connected. Each worker calls TcpTransport::worker with the coordinator’s address and its own rank.

Wire protocol

Every message is a length-prefixed f32 array:

1. 4-byte little-endian u32 element count.
2. count * 4 bytes of little-endian f32 values.

Rank coordination

Workers announce their rank as a 4-byte LE u32 immediately after the TCP handshake so the coordinator can place each connection in the correct slot.

Usage

For single-machine testing use loopback_pair which creates a coordinator + worker pair over 127.0.0.1 on a random port.

Implementations

impl TcpTransport

pub fn coordinator( bind_addr: &str, world_size: usize, ) -> Result<Self, ModelError>

Create a coordinator node that listens for worker connections.

Blocks until world_size - 1 workers have connected. The coordinator is always rank 0.

pub fn worker(coordinator_addr: &str, rank: usize) -> Result<Self, ModelError>

Create a worker node that connects to the coordinator.

pub fn send(&self, peer: usize, data: &[f32]) -> Result<(), ModelError>

Send a tensor’s data to a specific peer.

peer is an index into the peers list (0-based).

pub fn recv(&self, peer: usize) -> Result<Vec<f32>, ModelError>

Receive tensor data from a specific peer.

peer is an index into the peers list (0-based).

pub fn allreduce_sum(&self, data: &mut [f32]) -> Result<(), ModelError>

All-reduce: sum gradients across all nodes.

Implements a simple butterfly all-reduce pattern: the coordinator collects data from all workers, computes the element-wise sum, and broadcasts the result back.

pub fn rank(&self) -> usize

Returns the rank of this transport instance.

pub fn world_size(&self) -> usize

Returns the world size (total number of nodes).