Struct WorkerClient 

pub struct WorkerClient { /* private fields */ }

Worker client that connects to the coordinator.

Implementations

impl WorkerClient

pub fn connect( config: DistributedConfig, gpu_count: u32, backend: &str, ) -> Result<Self, String>

Connect to the coordinator and complete the join handshake.

Arguments

• config - Worker configuration with coordinator address
• gpu_count - Number of GPUs this worker has
• backend - Backend name (e.g., “wgpu”, “cuda”, “cpu”)

Errors

Returns error if connection or handshake fails.

pub fn receive_shard(&self) -> Result<Option<ShardAssignment>, String>

Receive shard assignment for the next training step.

Returns None if the coordinator sends a Shutdown message.

Errors

Returns error on communication failure.

pub fn send_gradients( &self, step: u64, gradients: Vec<f32>, loss: f32, correct: usize, total: usize, ) -> Result<(), String>

Send computed gradients to the coordinator.

Arguments

• step - Training step number
• gradients - Gradient vector (flattened LoRA params + classifier head)
• loss - Average loss for this shard
• correct - Number of correct predictions
• total - Total samples in shard

Errors

Returns error on send failure.

pub fn receive_averaged(&self) -> Result<AveragedResult, String>

Receive averaged gradients from coordinator after AllReduce.

Errors

Returns error on communication failure.

pub fn send_block_gradient( &self, step: u64, block_idx: u32, num_blocks: u32, gradients: Vec<f32>, component_sizes: Vec<u32>, ) -> Result<(), String>

Send per-block gradient to coordinator for AllReduce (v2 DDP).

Arguments

• step - Training step number
• block_idx - Transformer block index (0-based)
• num_blocks - Total number of transformer blocks
• gradients - Flattened gradient vector (9 components concatenated)
• component_sizes - Element count for each of the 9 components

pub fn receive_averaged_block(&self) -> Result<AveragedBlockResult, String>

Receive averaged block gradient from coordinator after AllReduce (v2 DDP).

pub fn send_non_block_gradient( &self, step: u64, component: u8, gradients: Vec<f32>, ) -> Result<(), String>

Send non-block gradient to coordinator for AllReduce (v2 DDP).

Arguments

• step - Training step number
• component - 0=lm_head, 1=final_norm, 2=embedding
• gradients - Gradient vector for this component

pub fn receive_averaged_non_block( &self, ) -> Result<AveragedNonBlockResult, String>

Receive averaged non-block gradient from coordinator after AllReduce (v2 DDP).

pub fn worker_id(&self) -> u32

This worker’s assigned ID

pub fn total_workers(&self) -> u32

Total number of workers in the cluster