Struct TensorGraphPartitioner 

pub struct TensorGraphPartitioner {
    pub nodes: HashMap<u64, NodeWeight>,
    pub edges: Vec<GraphEdge>,
}

Partitions a tensor computation graph into balanced subgraphs for distributed execution, minimising cross-partition edge traffic.

Fields

nodes: HashMap<u64, NodeWeight>

All nodes, keyed by their node_id.

edges: Vec<GraphEdge>

All directed edges in the graph.

Implementations

impl TensorGraphPartitioner

pub fn new() -> Self

Creates an empty TensorGraphPartitioner.

pub fn add_node(&mut self, node: NodeWeight)

Inserts (or replaces) a node in the graph.

pub fn add_edge(&mut self, edge: GraphEdge)

Appends an edge to the graph.

pub fn remove_node(&mut self, node_id: u64) -> bool

Removes a node and returns true if it was present.

Note: edges referencing the removed node are not automatically removed; callers that need edge consistency should manage this manually.

pub fn partition(&self, k: usize) -> Vec<Partition>

Partitions the graph into at most k parts.

Strategy

1. If k == 0 or the graph has no nodes, return an empty vector.
2. If k >= node count, assign one node per partition (only non-empty partitions are returned).
3. Otherwise, sort all nodes by compute_cost descending and assign each node to the partition with the lowest current total compute cost (greedy min-heap approximation via linear scan over k partitions).

pub fn stats(&self, partitions: &[Partition]) -> PartitionStats

Computes statistics for the given set of partitions.

Trait Implementations

impl Debug for TensorGraphPartitioner

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

Formats the value using the given formatter.

impl Default for TensorGraphPartitioner

fn default() -> TensorGraphPartitioner

Returns the “default value” for a type.