Struct ClusterRun 

pub struct ClusterRun {
    pub timings: Vec<StageTiming>,
    pub total_forward_ms: u64,
    pub output: Vec<f32>,
}

Full run result: the model output + per-node monitoring.

Fields

timings: Vec<StageTiming>

total_forward_ms: u64

output: Vec<f32>

Final-stage output (e.g. logits), flat.

Implementations

impl ClusterRun

pub fn tokens_per_s(&self, seq: usize) -> f64

Throughput in tokens/s given the sequence length driven.

pub fn reduce_resident_gb(&self, kind: ReduceKind) -> f64

Aggregate a per-node metric with a reduce op — the SAME ReduceKind semantics the collective all-reduce uses for tensor-parallel partial sums. (Here over locally-gathered stats; crate::cluster::all_reduce runs the real cross-node collective.)

pub fn critical_path_ms(&self) -> u64

Critical-path forward time = MAX over stages (a reduce-Max); the pipeline bottleneck the planner’s throughput policy minimizes.

pub fn table(&self) -> String

A monitor table for stdout.

Trait Implementations

impl Clone for ClusterRun

fn clone(&self) -> ClusterRun

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ClusterRun

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

Formats the value using the given formatter.