Expand description
Engine-neutral inference simulation, deterministic replay, and performance modeling.
engine owns scheduling, native GPU KV accounting, preemption, timing,
and attention-DP composition. replay owns virtual time, logical-worker
lifecycle, placement/scaling composition, and report collection.
perfmodel is the imported AIConfigurator latency and memory model. It is
deliberately namespaced so its EngineConfig and compiled engine do not
collide with the replay engine’s public API.
The crate root exposes the stable, commonly used configuration and replay surface. Advanced engine and adapter contracts remain available through their explicit module paths.
Re-exports§
pub use engine::EngineConfig as ReplayEngineConfig;pub use engine::TimingModel;pub use engine::TimingModelConfig;pub use replay::ReplayReport;pub use replay::ReplaySpec;pub use replay::Replayer;pub use perfmodel::EngineConfig;pub use perfmodel::AicError;pub use perfmodel::BackendKind;pub use perfmodel::DataType;pub use perfmodel::ENGINE_CONFIG_SCHEMA_VERSION;pub use perfmodel::ENGINE_SPEC_SCHEMA_VERSION;pub use perfmodel::EstimateSource;pub use perfmodel::FPM_VERSION;pub use perfmodel::ForwardPassMetrics;pub use perfmodel::ForwardPassPerfDiagnostics;pub use perfmodel::ForwardPassPerfModel;pub use perfmodel::ForwardPassPerfOptions;pub use perfmodel::ForwardPassPerfReadiness;pub use perfmodel::ForwardPassPerfSource;pub use perfmodel::KvCacheEstimate;pub use perfmodel::KvCacheEstimateAdjusted;pub use perfmodel::KvCacheEstimateError;pub use perfmodel::KvCacheEstimateOptions;pub use perfmodel::KvCacheEstimateRequest;pub use perfmodel::KvCacheMemoryFraction;pub use perfmodel::MemoryBreakdown;pub use perfmodel::ParallelMapping;pub use perfmodel::QuantizationConfig;pub use perfmodel::QueuedRequestMetrics;pub use perfmodel::ScheduledRequestMetrics;pub use perfmodel::SpeculativeConfig;