Skip to main content

CommunicationBackend

tensorlogic_infer::distributed

Trait CommunicationBackend 

Source 
pub trait CommunicationBackend: Send + Sync {

Show 13 methods    // Required methods
    fn initialize(
        &mut self,
        config: &DistributedConfig,
    ) -> Result<(), ExecutorError>;
    fn finalize(&mut self) -> Result<(), ExecutorError>;
    fn rank(&self) -> usize;
    fn world_size(&self) -> usize;
    fn all_reduce(
        &self,
        tensor_id: &str,
        reduction: ReductionOp,
    ) -> Result<(), ExecutorError>;
    fn broadcast(
        &self,
        tensor_id: &str,
        src_rank: usize,
    ) -> Result<(), ExecutorError>;
    fn scatter(
        &self,
        tensor_id: &str,
        src_rank: usize,
    ) -> Result<(), ExecutorError>;
    fn gather(
        &self,
        tensor_id: &str,
        dst_rank: usize,
    ) -> Result<(), ExecutorError>;
    fn all_gather(&self, tensor_id: &str) -> Result<(), ExecutorError>;
    fn reduce_scatter(
        &self,
        tensor_id: &str,
        reduction: ReductionOp,
    ) -> Result<(), ExecutorError>;
    fn send(
        &self,
        tensor_id: &str,
        dst_rank: usize,
    ) -> Result<(), ExecutorError>;
    fn recv(
        &self,
        tensor_id: &str,
        src_rank: usize,
    ) -> Result<(), ExecutorError>;
    fn barrier(&self) -> Result<(), ExecutorError>;

}
Expand description

Abstract communication backend for device-to-device communication.

Required Methods§

Source

fn initialize( &mut self, config: &DistributedConfig, ) -> Result<(), ExecutorError>

Initialize the communication backend.

Source

fn finalize(&mut self) -> Result<(), ExecutorError>

Finalize and clean up the backend.

Source

fn rank(&self) -> usize

Get the rank of this process.

Source

fn world_size(&self) -> usize

Get the world size (total number of processes).

Source

fn all_reduce( &self, tensor_id: &str, reduction: ReductionOp, ) -> Result<(), ExecutorError>

Perform an all-reduce operation.

Source

fn broadcast( &self, tensor_id: &str, src_rank: usize, ) -> Result<(), ExecutorError>

Broadcast from source rank to all ranks.

Source

fn scatter(&self, tensor_id: &str, src_rank: usize) -> Result<(), ExecutorError>

Scatter data from source rank to all ranks.

Source

fn gather(&self, tensor_id: &str, dst_rank: usize) -> Result<(), ExecutorError>

Gather data from all ranks to destination rank.

Source

fn all_gather(&self, tensor_id: &str) -> Result<(), ExecutorError>

All-gather operation.

Source

fn reduce_scatter( &self, tensor_id: &str, reduction: ReductionOp, ) -> Result<(), ExecutorError>

Reduce-scatter operation.

Source

fn send(&self, tensor_id: &str, dst_rank: usize) -> Result<(), ExecutorError>

Point-to-point send.

Source

fn recv(&self, tensor_id: &str, src_rank: usize) -> Result<(), ExecutorError>

Point-to-point receive.

Source

fn barrier(&self) -> Result<(), ExecutorError>

Synchronize all processes.

Implementors§