Struct DeviceMesh 

pub struct DeviceMesh { /* private fields */ }

An n-D arrangement of ranks. The product of shape must equal the world size; ranks are laid out in row-major order (last dim varies fastest).

Implementations

impl DeviceMesh

pub fn new(shape: Vec<usize>, world_size: usize) -> FerrotorchResult<Self>

Create a mesh with the given shape. world_size must equal the product of shape (verified eagerly so misconfigured launches fail loudly instead of silently splitting wrong).

pub fn new_with_names( shape: Vec<usize>, dim_names: Vec<String>, world_size: usize, ) -> FerrotorchResult<Self>

Variant of [new] that also attaches names to each dim. dim_names.len() must match shape.len().

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

Mesh shape ([dp, tp, ...]).

pub fn dim_names(&self) -> Option<&[String]>

Optional dim names.

pub fn ndim(&self) -> usize

Dimensionality of the mesh.

pub fn size(&self) -> usize

Total number of ranks in the mesh (= world_size).

pub fn dim_index(&self, name: &str) -> FerrotorchResult<usize>

Resolve a dim name to its index.

pub fn coords(&self, rank: usize) -> FerrotorchResult<Vec<usize>>

Convert a rank to its multi-dim coordinate within the mesh. Row-major: the last dim varies fastest.

pub fn rank_of(&self, coords: &[usize]) -> FerrotorchResult<usize>

Inverse of [coords]: convert a coordinate to its rank.

pub fn ranks_along_dim( &self, dim: usize, rank: usize, ) -> FerrotorchResult<Vec<usize>>

All ranks along dim that share rank’s coordinates on every other dim. Useful for constructing per-axis collective groups (e.g. one TP group per data-parallel slice).

Returns the ranks in increasing-coord order on dim.

pub fn groups_along_dim(&self, dim: usize) -> FerrotorchResult<Vec<Vec<usize>>>

All sub-groups along dim: a partitioning of the world into disjoint groups of shape[dim] ranks each, such that every group consists of ranks differing only on dim. Useful for building sub-backends in bulk.

Trait Implementations

impl Clone for DeviceMesh

fn clone(&self) -> DeviceMesh

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DeviceMesh

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.