Skip to main content

Crate rumus_distributed

Crate rumus_distributed 

Source
Expand description

rumus-distributed — 3D parallelism for RUMUS.

Tensor Parallelism (ColumnParallel + RowParallel), Pipeline Parallelism (1F1B micro-batch schedule), and async collective operations.

Re-exports§

pub use collective::AllReduceHandle;
pub use collective::CollectiveBarrier;
pub use collective::CommThread;
pub use pipeline::PipelineExecutor;
pub use pipeline::PipelineStage;
pub use tensor_parallel::ColumnParallelLinear;
pub use tensor_parallel::RowParallelLinear;

Modules§

collective
Async collective operations: AllReduce via dedicated comm threads.
pipeline
Pipeline parallelism: 1F1B micro-batch schedule with per-micro-batch tapes.
tensor_parallel
Tensor Parallelism: ColumnParallelLinear and RowParallelLinear.