Struct GpuBackendSession 

pub struct GpuBackendSession { /* private fields */ }

Opaque capability installed into session/optimizer state.

It cannot construct or expose raw runtime resources and its weak resource reference cannot keep device ownership alive after public facade close.

Implementations

impl GpuBackendSession

pub fn identity(&self) -> GpuBackendCoreIdentity

pub fn is_closing(&self) -> bool

Returns whether the backend close gate has been signalled.

An attempt that won the begin gate reads this flag and terminalizes unplanned with backend_closing evidence. Planning does not start in that case. The begin gate itself remains the only linearization point.

pub fn begin_attempt( &self, request: BeginAttemptRequest, ) -> Result<AttemptBegin>

Begins one attempt lifecycle under this backend’s observation scope.

The observation permit is the linearization point: an attempt that obtains it is counted in attempts_begun and holds the observation scope open until its report is ready, so close cannot complete without that report. A begin that loses to close fails with backend_closing and leaves no counted attempt. No GPU permit is taken here; the attempt claims a device only when its first native fragment is admitted.

pub fn backend_metrics_snapshot(&self) -> BackendMetricsSnapshot

Returns the same coherent generation snapshot exposed by the owning backend core without requiring access to live GPU resources.

pub fn take_abandoned_attempt_reports( &self, max: usize, ) -> Vec<Arc<QueryAttemptReport>>

Drains up to max reports of attempts whose completion receiver was dropped before the report was ready. Empty when the backend was built without an attempt completion queue.

pub fn attempt_counters_snapshot(&self) -> AttemptBackendCountersSnapshot

pub async fn admit_async_with_request( &self, request: QueryAdmissionRequest, ) -> Result<QueryHandle>

Admits one native attempt, waiting in the bounded admission queue until a grant is placed, the request is cancelled, or its queue deadline passes.

The observation and GPU permits are taken before the request enters the queue and travel with it inside its QueryCompletion, so an attempt that is queued, admitted, or cleaning up holds the resource generation for its whole lifetime, and close waits for it. On a rejected or cancelled request the GPU permit is released before the observation is sealed and delivered, the observation permit covers that delivery and is released afterwards, and the opaque activity lease runs last because its destructor may reenter close. Failures before the queue carry backend_closing (Cancelled); failures inside the queue are the admission controller’s (ServiceOverloaded for a full queue or an elapsed wait, the typed capability or health rejections otherwise).

pub fn same_backend(&self, other: &Self) -> bool

Whether both sessions were minted from the same published core.

pub fn diagnostics(&self) -> Result<BackendDiagnostics>

Reads every service report once. Like every service read, it fails with backend_closing once close begins.

pub fn device_telemetry_snapshots(&self) -> Result<Vec<DeviceTelemetrySnapshot>>

Trait Implementations

impl Clone for GpuBackendSession

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GpuBackendSession

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

Formats the value using the given formatter.