#![warn(missing_docs)]
pub mod automatic_support;
pub mod backend;
pub mod cache;
pub mod capture;
pub mod communication;
pub mod component;
pub mod composite;
pub mod decision;
pub mod dense;
pub mod draft;
pub mod execution;
pub mod execution_control;
pub mod execution_plan;
pub mod expert;
pub mod generation;
pub mod input;
pub mod inspection;
pub mod intervention;
pub mod layered;
pub mod load_request;
pub mod mechanism_synthesis;
pub mod parallel;
pub mod parameter;
pub mod partition;
pub mod partitioned_execution;
pub mod placement;
pub mod prefetch;
pub mod realtime;
pub mod realtime_executor;
pub mod realtime_ingress;
pub mod realtime_interpreter;
pub mod realtime_mechanism_synthesis;
pub mod realtime_model;
pub mod realtime_payload;
pub mod realtime_payload_state;
pub mod realtime_selection;
pub mod realtime_session;
pub mod replicated_session;
pub mod replicated_text;
pub mod residency;
pub mod speculative;
pub mod speculative_selection;
pub mod state;
mod weight_residency;
pub use automatic_support::{
residency_telemetry, selected_text_bounded_requirement, BoundedResidencySizingError,
};
pub use backend::{
BarrierBackend, BroadcastBackend, CollectiveBackend, CommunicationBackend, EvenGatherBackend,
FailureAgreementBackend, ParameterBackend, PointToPointBackend, RoleExactBoundaryValue,
SubmissionBackend, SumReductionBackend, TransferBackend, UnevenGatherBackend,
VariableAllToAllBackend,
};
pub use cache::{
finalize_prompt_cache_shard, hash_prompt_cache_shard_payload, inspect_prompt_cache,
prompt_cache_rank_path, resolve_prompt_cache_root, safe_prompt_cache_shard_path,
validate_prompt_cache_manifest, CacheBlockLifecycle, CacheBlockStorage,
CacheHostDemotionOperation, CacheHostPromotion, CacheIoAdmission, CacheIoCompletionDisposition,
CacheIoExecutionState, CacheIoExecutionStateError, CacheIoOperation, CacheIoOperationKey,
CacheIoOperationKind, CacheIoPreparation, CacheIoStartDisposition, CacheIoSubmission,
CacheIoSubmissionOutcome, CacheIoTicket, CacheIoWorker, CacheIoWorkerError,
CacheLayerResidencyReport, CacheLayerResidencyStats, CacheLifecycleError, CachePoolError,
CachePoolLimits, CachePoolMembership, CachePoolReport, CachePoolReservation, CachePoolResource,
CachePoolUsage, CacheResidencyConfigurationError, CacheResidencyPolicy, CacheResidencyPool,
CacheResidencyReport, CacheResidencyTelemetry, CacheStorageError, CacheStoragePhase,
LiveCacheBlockPublication, LiveCacheDiskPolicy, LiveCachePublicationError, MutableCacheTail,
PagedCacheOptions, PromptCachePersistenceError, PromptCachePublication,
ReversiblePromptCachePublication, CACHE_RESIDENCY_LAYER_REPORT_LIMIT,
MAX_PROMPT_CACHE_SHARD_HEADER_BYTES, PROMPT_CACHE_CURRENT_FILE,
PROMPT_CACHE_GENERATIONS_DIRECTORY,
};
pub use communication::validate_communication_manifest_consensus;
pub use communication::{
prepare_communication_realization, project_all_communication_manifests,
project_communication_manifest, validate_compatible_communication_manifests,
BoundaryDimensionContract, BoundaryFramingProtocol, BoundaryRoleContract,
CommunicationCapabilities, CommunicationCapabilityError, CommunicationCompletionCapabilities,
CommunicationCompletionPolicy, CommunicationGroupDescriptor, CommunicationGroupRequirements,
CommunicationManifest, CommunicationManifestConsensusError, CommunicationManifestError,
CommunicationOperation, CommunicationOperationRequirement, CommunicationPeerCounts,
CommunicationRealizationError, CommunicationRouteDescriptor, CommunicationRouteId,
CommunicationTensorLimits, CommunicationTopologyCapabilities, PreparedCommunicationRealization,
RoleExactBoundaryContract, TopologyCommunicationPlan,
};
pub use component::{
ComponentDomain, ComponentGraph, ComponentGraphError, ComponentKind, ComponentResidencyClass,
ComponentSpec,
};
pub use composite::{
select_composite_realization, select_processor_execution, CompositeSelectionError,
MediaPrimitiveCapabilities, ModalityProcessorRequirements, ProcessorExecutionRequirements,
ProcessorPrimitive, ProcessorSelectionError, ProcessorSelectionRequest,
SelectedCompositeRealization, SelectedProcessorExecution,
};
pub use decision::{
FullyForcedTailDecision, PredictionDirective, SequentialDecision, SequentialDecisionBoundary,
SequentialDecisionDiagnostic, SequentialDecisionDriver, SequentialDecisionError,
SequentialDecisionMode, SequentialDecisionPlan, SequentialDecisionPlanError,
SequentialDecisionSource, SequentialDecisionTraversal, SequentialSamplingState,
};
pub use dense::{
DenseCacheMetrics, DenseDiskStreamReport, DenseExecutionGroupReport, DensePassCounterSnapshot,
DensePassReport, DenseStreamTelemetry, DenseStreamTelemetryError, DenseTierResidencyReport,
};
pub use draft::{execute_draft_group, DraftGroupExecutionError, DraftStateTransaction};
pub use execution::{
ExecutionGraph, ExecutionGraphError, ExecutionGroupId, ExecutionGroupSchedule,
ExecutionGroupSpec, ExecutionScheduleError, ExecutionUnitAddress, ExecutionUnitLayout,
ExecutionUnitLayoutError, ReadyGroupState,
};
pub use execution_plan::{
execution_plan_quantization, ExecutionPlanLoadError, ResidencyDiagnostics,
};
pub use expert::{
combine_routed_expert_tensor_parallel, combine_tensor_parallel_expert_outputs,
plan_addressable_bank_bindings, reduce_routed_expert_tensor_parallel,
reduce_tensor_parallel_expert_output, selected_addressable_parameter_bytes,
AddressableBankBindingPlan, AddressableBankDistribution, AddressableBankMember,
AddressableBankMemberError, AddressableBankMemberPlacement, AddressableBankParameter,
AddressableBindingTransform, AddressableExpertRouteProvider, AddressableExpertRouteRequest,
AddressableGatedProductBank, AddressableGroupedBank, ExpertRouteCombination,
ExpertRouteExchange, ExpertRouteTensorMovement, IndexedMovement, ObservedExpertProvider,
ObservedExpertProviderError, ParameterBankAcquisition, ResidentExpertProvider,
RoutedExpertProvider, RoutedExpertRequest, RoutedExpertTensorParallelOutput,
RoutedObservationPoint, TensorParallelRoutedExpertProvider,
};
pub use expert::{select_routes_with_observer, select_routes_with_provider};
pub use generation::{
CausalModel, ConstrainedSampler, DefaultSampler, GenerationSampler, MirostatV2Sampler,
PenaltyConfig, Sampler, SamplingBackend, SamplingConfigurationError, SpeculativeSampler,
TokenDomain,
};
pub use input::{
PreparedInputCacheIdentity, PreparedInputCacheIdentityError, PreparedInputInspector,
PreparedInputPart, PreparedInputPayload, PreparedModelInput,
};
pub use inspection::{
observe_and_intervene, observe_model_logits, ActivationObserver, NoopObserver, RoutingDecision,
RoutingObservation, TargetStateCapture, TargetStateCaptureError, TargetStateTap,
};
pub use layered::{
ArchitectureGroupKind, ArchitectureGroupPlacement, ArchitectureGroupTransport,
ArchitectureMergeDestination, ArchitectureParallelSubgroup, ArchitectureParameters,
CompositeLayeredTraversalHook, LayeredArchitecture, LayeredForwardState, LayeredPartitionInput,
LayeredPartitionOutput, LayeredPipelineSchedule, LayeredPipelineScheduleError,
LayeredTraversalHook, LayeredTraversalPoint, LayeredUnitAction, LayerwiseAcquireError,
LayerwisePolicy, LayerwisePolicyForward, LayerwiseRuntime, LayerwiseRuntimeError,
ParallelLayeredArchitecture, ParallelRoutedLayeredArchitecture, PartitionedLayeredArchitecture,
ResidentRuntime, ResidentUnitWindow, ResidentUnitWindowError, RoutedLayeredArchitecture,
StaticParameterVisitor, StaticParameterVisitorMut,
};
pub use load_request::{
DraftingLoadRequest, NormalizedLoadRequest, NormalizedLoadRequestError, ParallelLoadRequest,
ValidatedModelLoadRequest,
};
pub use mechanism_synthesis::{
synthesize_replicated_text_capabilities, BackendMechanismFacts, ReplicatedTextMechanismSupport,
StateLifecycleCapabilities,
};
pub use parallel::{
aligned_partition_units, expand_linear_format_parameter_groups, module_parameter_group,
partitioned_module_parameter_group, partitioned_projection_group, projection_parameter_group,
segmented_projection_group, LocalModelLayout, LocalTensorLayout, MemberSharding,
ParallelModelInfo, ParallelPlanError, ParameterGroupSpec, ParameterMemberSpec, ParameterRole,
ProjectionSharding, ShardingPolicy, TensorPlacement,
};
pub use parameter::{
bind_materialized_unit, bind_materialized_unit_excluding, bindings_from_recipe_set,
build_exact_replicated_text_bindings, build_module_binding_plan, materialize_bindings,
materialize_selected_bindings, preflight_bindings, select_bindings, BindingPlan,
BindingPlanError, MaterializedUnit, ModuleBindingPlan, ModuleBindingPlanError,
ParameterBindingTarget, ParameterOrchestrationError, PlannedBinding, RecipeBindingError,
SelectedBindingPlan,
};
pub use partition::{
validate_boundary_tensor_count, ArchitectureBoundary, ArchitectureBoundaryError,
ArchitectureBoundaryValue, ArchitectureParameterDescription, ArchitectureParameterError,
ArchitecturePartition, ArchitecturePartitionError, BoundaryTensorDimension,
BoundaryTensorDtype, BoundaryTensorSpec, BoundaryWireSchema, LayeredPartitionBeginError,
LayeredPartitionDriver, LayeredPartitionError, NoAuxiliaryBoundary, NoAuxiliaryBoundarySchema,
OwnedParameterGroupSpec, ParameterGroupOwner, PartitionGroup, PartitionOwnership,
PartitionState, PipelineActivationDtype, PipelineWireContract, ResolvedBoundaryTensorSpec,
ResolvedBoundaryWireSchema,
};
pub use partitioned_execution::{
CommunicationTensorMetadata, DistributedExecutionPhase, LayerwiseTraversalPartitionExecutor,
LayerwiseTraversalRuntime, NoBoundaryTransport, NoCommitAgreement, NoOutputPublisher,
OpaqueBoundaryTransport, OpaqueCommitAgreement, OpaqueFailureAgreement, OpaqueOutputPublisher,
PartitionBoundaryRoute, PartitionBoundaryTransport, PartitionCommitAgreement,
PartitionCommunication, PartitionCommunicationAuthority, PartitionExecutionError,
PartitionOutputAuthority, PartitionOutputPublication, PartitionOutputPublisher,
PartitionedExecutionPlan, PartitionedGroupExecutor, PartitionedTextExecution,
PartitionedTextRuntime, PartitionedTraversalError, PartitionedTraversalResult,
RealizedCommunicationGroup, RealizedCommunicationRoute,
};
pub use placement::{
place_addressable_member_bindings, place_weight_bindings, placement_selection,
BindingPlacementError, PlacementPlan, PlacementPlanError, PlacementRank,
ResolvedTensorPlacement, TensorSlice,
};
pub use prefetch::{BackgroundPrefetchWorker, BackgroundPrefetchWorkerError};
pub use realtime::{
RealtimeCompletionAttachmentError, RealtimeFrameExecutionError, RealtimeFrameTransition,
RealtimeGenerationBranch, RealtimeGenerationState, RealtimeGenerationTransactionError,
};
pub use realtime_executor::{
execute_realtime_frame, PreparedRealtimeFrameExecutor, PrepublicationRealtimeFrame,
RealtimeCompletionCreationError, RealtimeDecisionExecution, RealtimeFrameCompletionMechanism,
RealtimeFrameCoordinatorError, RealtimeFrameHostObserver, RealtimeHostOutputUnavailable,
RealtimePrepublicationError, SubmittedRealtimeFrame,
};
pub use realtime_ingress::{
MaterializedRealtimeInput, RealtimeHostTokenMaterializer, RealtimeIngressContract,
RealtimeIngressError, RealtimePayloadKind, RealtimeTokenKind, ValidatedRealtimeInput,
};
pub use realtime_interpreter::{
complete_realtime_frame, prepare_realtime_frame, CompletedRealtimeFrame, PreparedRealtimeFrame,
RealtimeFrameInterpretationError, RealtimeFrameTensorMechanisms,
};
pub use realtime_mechanism_synthesis::{
synthesize_realtime_capabilities, RealtimeMechanismFacts, RealtimeMechanismSupport,
};
pub use realtime_model::{
construct_realtime_model, preflight_realtime_materialization_tasks, realtime_task_binding_plan,
ConstructedRealtimeExecution, ConstructedRealtimeModel, PreparedRealtimeModelContract,
RealizedRealtimePolicy, RealizedRealtimeState, RealtimeArchitectureConstructionIdentity,
RealtimeArchitectureIdentity, RealtimeLayerwiseRuntime, RealtimeMaterializationComponent,
RealtimeMaterializationTask, RealtimeModelConstructionError,
RealtimeModelConstructionMechanisms, RealtimeModelContractError, RealtimeTaskBindingPlan,
};
pub use realtime_payload::{
RealtimePayloadContract, RealtimePayloadContractError, RealtimePayloadEnvelope,
RealtimePayloadGeneration, RealtimePayloadHistory, RealtimePayloadHistoryError,
RealtimePayloadOwnerIdentity,
};
pub use realtime_payload_state::{
RealtimePayloadBranch, RealtimePayloadState, RealtimePayloadStateTransactionError,
};
pub use realtime_selection::{
select_and_prepare_realtime_realization, select_realtime_realization,
ConstructedRealtimeResources, PreparedRealtimeRealization, RealtimeArchitectureProof,
RealtimeArchitectureRequirements, RealtimeContractError, RealtimeExecutionRequirements,
RealtimeIdentity, RealtimeMechanism, RealtimeMechanismCapabilities,
RealtimeMechanismRequirements, RealtimeObservationRequirements, RealtimePreparationError,
RealtimeSelectionError, RealtimeSelectionIssue, RealtimeSelectionRequest,
RealtimeTopologyPolicy, RealtimeWeightComponentRequirement, RealtimeWeightComponentRole,
RealtimeWeightLoweringRequirement, SelectedRealtimeRealization,
SelectedRealtimeStateComponentRealization, SelectedRealtimeStateRealization,
};
pub use realtime_session::{
RealtimeHistoryGeneration, RealtimeModelOwnerIdentity, RealtimeModelSessionIdentity,
RealtimeSamplingReplacementError, RealtimeSamplingUpdateError, RealtimeSessionBranch,
RealtimeSessionError, RealtimeSessionExecutionError, RealtimeSessionIncarnation,
RealtimeSessionResumeError, RealtimeSessionScheduler, RealtimeSessionState,
RealtimeSessionTransactionError, ReleasedRealtimeSession,
};
pub use replicated_session::{
construct_replicated_text_session, construct_replicated_text_session_with_execution,
construct_replicated_text_session_with_runtime, prepare_default_partitioned_runtime,
prepare_layered_text_contract, prepare_layered_text_contract_with_addressable_parameters,
prepare_partitioned_session_runtime, prepare_partitioned_session_runtime_with_exclusions,
prepare_replicated_text_contract, prepare_replicated_text_contract_with_addressable_parameters,
DirectReplicatedTextExecution, DistributedSessionCheckpoint, DistributedStateCheckpoint,
PartitionedRuntimeConstructionError, PartitionedSessionFactoryInput,
PartitionedSessionPreparationError, PartitionedUnitScope, PredictionTargetOperation,
PreparedPartitionedRuntimeComponents, PreparedPartitionedSessionRuntime,
PreparedReplicatedTextContract, ReplicatedRuntimeExecutionStrategy,
ReplicatedTextExecutionStrategy, ReplicatedTextRuntime, ReplicatedTextSession,
ReplicatedTextSessionCheckpoint, ReplicatedTextSessionError, ReplicatedTextSessionMechanisms,
ReplicatedTextSessionReport, RoutedReplicatedTextExecution, SessionStateRealization,
TransactionalPromptCacheMechanisms,
};
pub use replicated_text::{
group_replicated_text_transform_tasks, locally_materialized_replicated_text_outputs,
partition_selected_replicated_text_materialization_tasks,
partitioned_replicated_text_materialization_tasks,
plan_local_replicated_text_materialization_tasks, plan_replicated_text_materialization_tasks,
replicated_text_materialization_tasks, select_replicated_text_realization,
selected_materialization_task_bytes, AddressableStorageCapabilities, AddressableStorageTiers,
BackendMechanismCapabilities, GroupedOperationRequirement, ParameterTransformConstraint,
ParameterTransformTarget, ReplicatedTextArchitecture, ReplicatedTextContractError,
ReplicatedTextMaterializationPartitionPlan, ReplicatedTextMaterializationTask,
ReplicatedTextOutputCompanion, ReplicatedTextOutputSelection, ReplicatedTextParameterOwner,
ReplicatedTextParameterPresence, ReplicatedTextParameterRequirement,
ReplicatedTextParameterRole, ReplicatedTextPhysicalSource, ReplicatedTextRequirements,
ReplicatedTextSelectionError, ReplicatedTextSelectionRequest, ReplicatedTextStateAccess,
ReplicatedTextTransformGroup, SelectedParameterRealization, SelectedReplicatedTextRealization,
SelectedStateComponentRealization, SelectedStateRealization, StateComponentMechanism,
StateComponentPlacement, StateMechanismCapabilities, StateStorageDtype,
WeightLoweringCapability, WeightLoweringDescriptor, WeightLoweringKind,
WeightResidencyMechanism,
};
pub use residency::{
DeviceLayerWindow, OffloadUnit, QuantizationCompanionBindings, ResidencyAcquisition,
ResidencyController, ResidencyControllerError, ResidencyDeclarationError, ResidencyLease,
ResidencyLeaseOwner, ResidencyLeaseStorage, ResidencyReport, ResidencyTransfer,
ResidencyTransferOwner, ResidencyWindowError, ResidencyWindowManager, ResidentLayerGroup,
ResidentLayerGroupReport, WeightBinding, WeightBindingPlan, WeightBindingSelectionError,
WeightMaterializationReport,
};
pub use speculative::{RunSpeculativeGeneration, SpeculativeScheduler};
pub use speculative_selection::{
select_and_prepare_speculative_realization,
select_and_prepare_speculative_realization_observed, select_speculative_realization,
ConstructedSpeculativeResources, PreparedSpeculativeRealization, SelectedSpeculativeCompletion,
SelectedSpeculativePlacement, SelectedSpeculativeRealization, SelectedSpeculativeSampling,
SelectedSpeculativeState, SpeculativeArchitectureCompatibilityProof, SpeculativeCaptureEntry,
SpeculativeCaptureEnvelope, SpeculativeCaptureError, SpeculativeCaptureMetadata,
SpeculativeCaptureSchema, SpeculativeContractError, SpeculativeIdentity,
SpeculativeLaneIdentity, SpeculativeMechanism, SpeculativeMechanismCapabilities,
SpeculativeMechanismRequirements, SpeculativePlacementRequest, SpeculativePreparationError,
SpeculativeRealizationRequirements, SpeculativeSelectionError, SpeculativeSelectionRequest,
SpeculativeStateCacheIdentityIngredients, SpeculativeStrategyClass,
SpeculativeStrategyRequirements,
};
pub use state::{
realize_architecture_state, ArchitectureStateFactory, ArchitectureStatePartitionError,
ArchitectureStatePartitionPlan, ArchitectureStatePartitionRule, ArchitectureStatePlacement,
ArchitectureStateRealizationError, DeviceState, LayerRuntimeState, ModelStateIdentity,
ResettableRuntimeLayerState, ResettableRuntimeState, RuntimeLayerState, RuntimeState,
RuntimeStateComponents, StateError, StateLayout, StateSegmentId, StateSegmentLifetime,
StateSegmentSpec, DEFAULT_STATE_SEGMENT_ID,
};
pub use weight_residency::{
DenseDiskStreamLoadOptions, DenseTransferSchedule, DenseTransferScheduleError,
ExecutionResidency, ExpertPass, LayerWeightResidency, LayerwiseLoadOptions,
LayerwiseModelMetadata, OrdinaryWeightResidency, ParameterBankAccess, ParameterBankKey,
ParameterBankLoadOptions, ParameterBankResidency, StaticUnitBindings, WeightResidency,
WeightResidencyPolicyError, DENSE_TRANSFER_WINDOW,
};