Struct ExecutorExecutionCapacityDeferral

Source

pub struct ExecutorExecutionCapacityDeferral { /* private fields */ }

Implementations§

Source §

impl ExecutorExecutionCapacityDeferral

Source

pub fn from_backing_pressure( observed: ExecutorAdmissionEpochs, wait_condition: CapacityWaitCondition, pressure: DynamicBackingPressure, stage: ExecutorExecutionCapacityStage, ) -> Result<Self>

Construct a physical deferral for an executor that observes device or pool pressure directly rather than through Ferrum’s backing allocator.

Source

pub fn from_admission( deferred: &AdmissionDeferred, stage: ExecutorExecutionCapacityStage, ) -> Result<Self>

Source

pub fn from_pending_maintenance( deferred: &AdmissionDeferred, stage: ExecutorExecutionCapacityStage, ) -> Result<Self>

Export an unresolved logical backing-growth decision after the executor’s bounded in-call maintenance attempts are exhausted.

This remains a pre-submit scheduling deferral. The bound controls only how much allocator maintenance one executor call may perform; it must not turn temporary capacity pressure into a terminal request failure.

Source

pub fn from_backing( deferred: &DynamicBackingDeferred, stage: ExecutorExecutionCapacityStage, ) -> Result<Self>

Export unresolved physical backing pressure without discarding its exact pool/domain evidence.

Source

pub fn with_relevant_maintenance_retry( self, attempts: u32, receipts: &[DynamicPoolGrowthBatchReceipt], pools: &[DynamicPoolStatus], affected_request_ids: Vec<RequestId>, ) -> Result<Self>

Attach a scheduler retry only when this call produced at least one real physical mutation relevant to the final blocker. An empty or unrelated receipt set is an ordinary typed wait, not an internal error.

Source

pub fn from_admission_maintenance( source: &AdmissionDeferred, observed: ExecutorAdmissionEpochs, wait_condition: CapacityWaitCondition, pressure: DynamicBackingPressure, maintenance_boundary: Option<DynamicPoolMaintenanceBoundaryReceipt>, stage: ExecutorExecutionCapacityStage, ) -> Result<Self>

Source

pub fn from_backing_maintenance( source: &DynamicBackingDeferred, observed: ExecutorAdmissionEpochs, wait_condition: CapacityWaitCondition, pressure: DynamicBackingPressure, maintenance_boundary: Option<DynamicPoolMaintenanceBoundaryReceipt>, stage: ExecutorExecutionCapacityStage, ) -> Result<Self>

Source

pub const fn maintenance_boundary( &self, ) -> Option<&DynamicPoolMaintenanceBoundaryReceipt>

Source

pub fn maintenance_retry(&self) -> Option<&ExecutorExecutionMaintenanceRetry>

Source

pub fn validated_maintenance_retry_scope( &self, current_request_ids: &[RequestId], ) -> Result<Option<&ExecutorExecutionMaintenanceRetry>>

Validate the bound retry scope against the authoritative logical frontiers in the current engine call.

Source

pub fn narrower_prefill_tokens(&self, attempted_tokens: usize) -> Option<usize>

Return a strictly smaller, capacity-informed prefill width.

This is a cold pressure-path hint, not allocator authority. The caller must probe the returned prefix through normal typed admission before any provider encode or device submission. A bounded reduction prevents a large frontier from producing an unbounded sequence of near-identical probes when the shortfall is small.