Struct AshaPruner 

pub struct AshaPruner { /* private fields */ }

The asynchronous successive-halving pruner.

The promotion rule, precisely

On a report at step:

1. If step is not exactly a rung resource level, the trial has not just reached a rung — Decision::Continue. (Deciding only at exact rung steps is what lets the value being judged, and every competitor’s value, be the value at that resource; a trial whose cadence never lands on a rung is never gated there, a safe omission — it is never wrongly pruned.)
2. Otherwise gather the values of every trial that reached this rung — from the typed StudyView, not a side-table — and apply is_promotable: promoted iff in the top 1/η, with the first η − 1 arrivals promoted optimistically so a rung baseline can form. Promote → Decision::Continue; otherwise Decision::Prune.

A trial pruned at a rung keeps its value there as its objective value (the pruned-adopts-last-intermediate rule — why a pruning example’s curve falls), which is exactly what makes the rung comparison for later trials sound.

Multi-objective

Successive halving ranks a single objective, so a multi-objective study is never pruned (returns Decision::Continue) rather than gate on one dimension and risk pruning a Pareto-optimal trial.

Determinism

The decision is a pure function of the ask-time StudyView, so it depends on which trials had filled the rung when this one was asked — completion-order dependent under parallelism. An ASHA-scheduled study is therefore replayable, not pre-determined (where determinism stops); under a single worker with a fixed seed the pruning is fully reproducible.

Where the rung table lives

Nowhere persistent. Optuna records rung membership as system_attrs strings; atune derives it from the StudyView’s typed intermediate reports on each call. The typed shape it takes when a caller wants to store or inspect it is RungTable, available through rung_table and round-trippable through a StateBlob on Scope::Scheduler — typed state, never attr-string hacks. Because the decision needs no persisted state, state returns None.

use atune_core::scheduler::AshaPruner;

// Rungs at 1, 3, 9 env-steps (η = 3), finishing at 27.
let asha = AshaPruner::new(1, 27, 3)?;
assert_eq!(asha.ladder().rungs(), [1, 3, 9]);

Implementations

impl AshaPruner

pub fn new( min_resource: u64, max_resource: u64, reduction_factor: u32, ) -> Result<Self>

Builds an ASHA pruner over the ladder (min_resource, max_resource, reduction_factor).

Errors

Error::InvalidSpace for a nonsense ladder — see RungLadder::new.

pub fn with_default_eta(min_resource: u64, max_resource: u64) -> Result<Self>

Builds an ASHA pruner with the default reduction factor.

Errors

As new.

pub const fn from_ladder(ladder: RungLadder) -> Self

Builds an ASHA pruner from an already-constructed ladder.

pub const fn with_objective(self, obj: usize) -> Self

Selects which objective dimension to gate on (default 0).

pub const fn ladder(&self) -> &RungLadder

The resource ladder this pruner gates on.

pub fn rung_table(&self, view: &StudyView) -> RungTable

The typed rung table derived from view — the occupants of every rung.

The Scope::Scheduler state shape the design mandates, built on demand rather than persisted: it round-trips through a StateBlob and is handy for a dashboard, a test, or M4’s DEHB.

Trait Implementations

impl Clone for AshaPruner

fn clone(&self) -> AshaPruner

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AshaPruner

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

Formats the value using the given formatter.

impl Scheduler for AshaPruner

fn on_report( &self, study: &StudyView, _trial: &TrialMeta, step: u64, values: &[f64], ) -> Result<Decision>

Promotes or prunes at each rung; see the type docs.

Errors

Never — the decision is total.

fn state(&self) -> Result<Option<SchedulerState>>

None: the rung table is derived from the StudyView, not persisted (see the type docs).

fn fan_report_mode(&self) -> FanReportMode

Declares whether this scheduler accepts aligned aggregate fan reports.

fn scripted_capability(&self) -> ScriptedSchedulerCapability

Declares whether scripted ask/tell can recreate this scheduler.

fn on_trial_end( &self, study: &StudyView, trial: &FrozenTrial, ) -> Result<Vec<Command>>

Called once a trial reaches a terminal state.

fn resume_candidates(&self, study: &StudyView) -> Result<Vec<Command>>

Paused trials worth waking on a freed work slot, in priority order.

fn restore_state(&self, blob: &SchedulerState) -> Result<()>

Restores a scheduler from a previously persisted blob.