Struct DTensor 

pub struct DTensor<T: Float> { /* private fields */ }

A logical tensor distributed across a DeviceMesh.

placements.len() must equal mesh.ndim() — there’s exactly one placement per mesh dim. The physical storage is the per-rank local_tensor; global_shape is the logical shape callers see.

Implementations

impl<T: Float> DTensor<T>

pub fn from_local( local_tensor: Tensor<T>, mesh: DeviceMesh, placements: Vec<Placement>, global_shape: Vec<usize>, ) -> FerrotorchResult<Self>

Wrap a per-rank local tensor with explicit placement annotations.

placements.len() must equal mesh.ndim(). global_shape is the logical full-tensor shape (for Replicate it equals local_tensor.shape(); for Shard it’s the local shape with the sharded dim multiplied by the mesh size along that dim).

pub fn from_local_replicated( local: Tensor<T>, mesh: DeviceMesh, ) -> FerrotorchResult<Self>

Build a fully-replicated DTensor: every rank holds the same tensor. Equivalent to from_local with placements = [Replicate; mesh.ndim()] and global_shape = local_tensor.shape().

pub fn to_local(&self) -> &Tensor<T>

The local shard held by this rank.

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

Logical full-tensor shape across all ranks.

pub fn placements(&self) -> &[Placement]

Per-mesh-dim placement annotations.

pub fn mesh(&self) -> &DeviceMesh

The associated mesh.

pub fn numel(&self) -> usize

Logical numel (product(global_shape)).

pub fn redistribute( &mut self, target_placements: Vec<Placement>, ) -> FerrotorchResult<()>

Redistribute this DTensor to a new placement spec.

target_placements.len() must equal mesh.ndim(). The supported transitions are:

• Replicate → Replicate: no-op.
• Shard(d) → Shard(d) (same dim): no-op.
• Replicate → Shard(d): scatter (caller picks the rank’s shard).
• Shard(d) → Replicate: all_gather along the relevant mesh dim.
• Partial(op) → Replicate: all_reduce along the relevant mesh dim.
• Shard(d) → Shard(e) with d != e: all_to_all transpose.

The actual collective dispatch is delegated to the crate::collective surface; this method records the intended target and updates local_tensor to reflect the local shard after the redistribute. Each transition is only validated for shape compatibility here — the cross-rank communication is performed by the lower-level crate::collective::* ops the caller invokes before / after redistribute lands. (This separation keeps the DTensor API testable without real multi-process launches.)

Trait Implementations

impl<T: Clone + Float> Clone for DTensor<T>

fn clone(&self) -> DTensor<T>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<T: Debug + Float> Debug for DTensor<T>

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

Formats the value using the given formatter.