Module perfmodel 

Rust-native AIConfigurator performance-model implementation.

The compiled-engine path is the only supported entry point: Python’s compile_engine walks the model once and emits an engine::spec::EngineSpec (op lists + EngineConfig identity); the Rust engine::Engine executes it without re-entering Python. With the python feature enabled, [AicEngineBuilder] is the preferred Rust → Python → Rust embedded build entry point and [AicEngine] is the PyO3 hot-path pyclass.

This directory remains a stable mirror of the former AIConfigurator Rust crate. Keeping the imported implementation behind one namespace makes upstream syncs mechanical while the crate root remains owned by replay.

Re-exports

pub use memory::EstimateSource;

pub use memory::KvCacheEstimate;

pub use memory::KvCacheEstimateAdjusted;

pub use memory::KvCacheEstimateError;

pub use memory::KvCacheEstimateOptions;

pub use memory::KvCacheEstimateRequest;

pub use memory::KvCacheMemoryFraction;

pub use memory::MemoryBreakdown;

Modules

engine

Compiled-engine layer.

memory

KV-cache memory estimation.

Structs

EngineConfig

Static engine identity and setup information carried by an crate::perfmodel::engine::spec::EngineSpec.

ForwardPassMetrics

Per-iteration forward-pass metrics.

ForwardPassPerfDiagnostics

Current readiness and tuning state for a ForwardPassPerfModel.

ForwardPassPerfModel

Forward-pass-level performance model with optional online tuning.

ForwardPassPerfOptions

In-memory tuning controls for ForwardPassPerfModel.

ParallelMapping

Parallelism layout. Flattened into EngineConfig so the flat wire keys (tp_size, pp_size, …) parse unchanged.

QuantizationConfig

Precision/quantization dtypes. Flattened into EngineConfig. Field names and types are unchanged from the former flat struct so the flat wire keys (weight_dtype, moe_dtype, …) parse unchanged.

QueuedRequestMetrics

Metrics for requests queued but not scheduled in one iteration.

ScheduledRequestMetrics

Metrics for requests scheduled in one forward-pass iteration.

SpeculativeConfig

Multi-Token Prediction speculative-decoding parameters. Wrapped in Option<> on EngineConfig so models without MTP don’t carry the noise, and #[serde(flatten)]-ed so the flat wire key (nextn) parses unchanged. Accepted-token progress is modeled above aic-core.

Enums

AicError

All errors surfaced by the Rust core.

BackendKind

Backend performance database family.

DataType

Precision/quantization dtypes carried on the engine-config wire.

ForwardPassPerfReadiness

Readiness state reported by ForwardPassPerfDiagnostics.

ForwardPassPerfSource

Prediction backend currently used by ForwardPassPerfModel.

Constants

ENGINE_CONFIG_SCHEMA_VERSION

ENGINE_SPEC_SCHEMA_VERSION

FPM_VERSION

Wire-schema version for ForwardPassMetrics (Dynamo’s scheduler telemetry). The FPM input is rejected with AicError::UnsupportedSchemaVersion when version does not match.