Expand description
Backend-neutral automatic execution planning. Backend-neutral automatic execution planning.
Backends report hardware, artifact resources, and candidate admission. This module owns policy validation, resource budgeting, plan selection, feedback matching, and the serialized planning and telemetry documents.
Structs§
- Allocator
Telemetry - Backend allocator observations for one execution.
- Automatic
Plan Request - Owned input to one automatic planning session.
- Automatic
Planner - Backend-neutral automatic planner.
- Automatic
Planner Policy - Tunable, serializable automatic-planning policy.
- Bounded
Residency Requirement - Exact bounded device-window requirement established by a backend probe.
- Candidate
Admission - Backend candidate-admission result consumed by the neutral planner.
- Duration
Seconds - Floating-point duration wrapper with equality based on its bit pattern.
- Execution
Plan Report - Complete automatic-planning document suitable for JSON persistence.
- Execution
Plan Target - One target-backend instance and retained selection realized from a portable execution plan.
- Execution
Plan Target Selection - Backend-owned preparation selected before native target realization.
- Execution
Telemetry - Stable JSON telemetry for one completed model execution.
- Expert
Cache Telemetry - Routed-expert cache occupancy summary.
- External
Draft Artifact - Architecture-prepared assistant artifact and proven portable tokenizer compatibility.
- Hardware
Backend Profile - Availability and devices for one execution backend.
- Hardware
Device Profile - One logical device visible to an execution backend.
- Hardware
Profile - Hardware and memory observations used as automatic-planning inputs.
- Model
Resource Profile - Architecture and header-derived planning facts used before a model is loaded.
- Plan
Explanation - Explanation accompanying a selected execution plan.
- Plan
Explanation Entry - One stable, machine-routable planner explanation entry.
- Residency
Telemetry - Logical bytes and transfers reported by bounded parameter residency.
- Retained
Automatic Plan - A portable report paired with the one artifact inspection used by every probe.
- Selected
Execution Plan Drafting - An external-drafting selection inseparably bound to its complete execution plan.
- Selected
Execution Plan Target - An inspected artifact and authoritative backend selection ready for native realization.
- Speculative
Decoding Telemetry - Speculative-decoding observations for one request.
- Timing
Telemetry - Timings reported for one generation request.
- Tokenizer
Compatibility Error - A target and external assistant do not share the same token-id vocabulary mapping.
- Tokenizer
Compatibility Proof - Proof that a target and external assistant use the same token-id vocabulary mapping.
- Transfer
Telemetry - One logical residency transfer counter.
Enums§
- Automatic
Planning Error - Failure produced by portable planning or its selected backend adapter.
- Hardware
Memory Semantics - Serializable form of physical host/device memory semantics.
- Observation
Kind - Confidence attached to an observed or derived value.
- Observed
- A value which remains explicit when the runtime cannot produce it.
- Plan
Explanation Level - Severity of one planner explanation entry.
- Realized
Drafting - Backend-owned drafting resources realized for one complete execution plan.
Constants§
- AUTOMATIC_
SCHEMA_ VERSION - Schema version shared by automatic-planning and telemetry documents.
Traits§
- Automatic
Planning Backend - High-level observations a backend supplies to the neutral planner.
- Execution
Plan Backend Factory - Selects and creates an executable whole-model backend from a portable execution plan.
Functions§
- realize_
execution_ plan_ drafting - Validates and realizes the drafting portion of a portable execution plan.
- realize_
execution_ plan_ target - Realizes native target resources for an already validated selection.
- select_
execution_ plan_ drafting - Selects drafting materialization before any native target realization.
- select_
execution_ plan_ target - Validates and selects the target portion before any native realization.
- speculative_
decoding_ telemetry - Projects neutral speculative statistics into the stable telemetry document.
Type Aliases§
- Execution
Plan Target Load Error - Failure while preparing the model retained by an execution-plan target.
- Prepared
Execution Plan Target - A realized backend paired with the model prepared from its retained selection.