Skip to main content Crate inferlab_protocol Copy item path Source AdapterError A structured rejection an integration returns in an AdapterResponse::Error AdapterProtocol The schema root for the independently released framework adapter SDK. AuxiliaryModelInput A logical auxiliary weight identity supplied during serving planning and
rendering: its governed kind plus the referenced logical model
([[RFC-0003:C-SERVE-AUXILIARY-MODELS]]). AuxiliaryModelLocator A machine-resolved auxiliary weight locator on a model-rank allocation
([[RFC-0003:C-SERVE-AUXILIARY-MODELS]]). BenchAgenticBranchStats BenchAgenticCatalogInput BenchAgenticResultEvidence BenchAgenticRunEvidence BenchAgenticSourceInput One release-qualified agentic trace source and its complete effective policy. BenchAgenticSourceVerification BenchCaseInput A single Bench case: its load shape and the number of requests to send. BenchClientRequest The request the Bench measurement runtime passes to its client: the
endpoint, model, bench definition, the load case to run, and the artifact
directory. BenchClientResult The result a Bench client writes for the measurement runtime to consume. BenchCorpusInput One operator-supplied text corpus binding for the random request source;
entry content is drawn as exact token-length slices of the corpus token
stream ([[RFC-0004:C-BENCH-REQUEST-SOURCES]]). BenchDatasetCatalogInput Immutable release-catalog facts resolved by the Rust control plane. BenchDatasetFilterInput BenchDefinitionInput The workload shape a Bench client drives, shared across its load cases. BenchInclusiveUniformInput The one bounded distribution currently supported for a token selector. BenchNativeInvocation One bounded native post-processing command and its terminal outcome. BenchPopulationInput One frozen dataset population consumed sequentially by every Bench case. BenchPopulationPreparationRequest The bounded tokenizer-backed operation that freezes one request population
before any Bench case starts. BenchPopulationPreparationResult Terminal result of request-population materialization. BenchPrefixConditioningInput File-bound exact canonical prefix used to condition a primed cache start. BenchPrefixGeometrySummary Resolved exact final-prompt prefix geometry across one frozen population. BenchPromptCacheObservation Backend-reported prompt/cache token accounting for one completed profiling request. BenchPromptTemplateProjection The concrete template used only for local complete-prompt projection. BenchPromptTokenReconciliation One synthetic profiling request’s planned-to-observed prompt-token check. BenchPromptTokenTargetingSummary Summary of best-effort complete-prompt targeting for a synthetic population. BenchRandomShapeInput One exact ISL/OSL pair and its relative categorical sampling weight. BenchRequestSloInput Per-request latency bounds lowered to the release-owned Bench runner. BenchRequestSloResult File-bound request-SLO evidence derived from AIPerf profiling records. BenchRuntimeSessionResult Terminal evidence for one admitted runtime session. BenchSessionDatasetCatalogInput Immutable release-catalog facts for a linear-session materializer. AIPerf
loader names are deliberately absent from this source boundary. BenchSessionPhaseSummary Reconciled native evidence for one session-bounded Bench phase. BenchSessionResultEvidence Session and transport reconciliation returned by the Bench client. BenchSessionSourceInput One release-qualified population of dependent linear session templates. BenchSessionTemplateInput One frozen linear-template summary needed to assign case slices without
duplicating the content-bearing population evidence artifact. BenchSessionTurnResult One native transport request reconciled to a linear-session turn. BenchSharedSystemContentSummary Resolved pre-template shared system-content shape across one frozen
server-chat population. BenchTokenCountSummary Summary of the realized token counts. Exact per-entry values remain in the
population evidence artifact. CaptureTargetRequirement Marks a replica as a profiling capture target and carries its window
control ([[RFC-0004:C-WORKLOAD-PROFILING]]). CaptureWindowControlRequirement The logical workload endpoint and typed actions that open and close a
capture window. CaptureWindowHttpActionSpec A capture-window HTTP action invoked against a logical serving endpoint.
The integration owns any framework-specific JSON body; the control plane
owns execution and evidence. ClientEndpointInput The public workload endpoint an Eval or Bench client connects to. EndpointAssignment A concrete host/port endpoint the control plane allocated for a process. EndpointDeclaration The plan-response endpoint capability declaration
([[RFC-0006:C-OPENAI-ENDPOINT-CONTRACT]]): the protocol plus the
integration-owned capability members (server metrics, prefix-cache
actions, and the cache-read representation). The control plane owns the
route paths, so this declaration carries no path values and rejects them
as unknown fields. EndpointRequirement The resolved workload endpoint requirement: the accepted declaration plus
the named OpenAI paths. The control plane fills the pinned
/v1/completions and /v1/chat/completions paths at plan acceptance
([[RFC-0006:C-OPENAI-ENDPOINT-CONTRACT]]); resolved evidence and client
inputs preserve both strings exactly. EvalClientRequest The request the Eval measurement runtime passes to its client: the endpoint
to hit, the model, the eval definition, and where to write artifacts. EvalClientResult The result an Eval client writes for the measurement runtime to consume. EvalMetricGate The effective threshold comparison for the configured primary metric. EvalNormalizedMetric One finite lm-eval metric with its exact native provenance. EvalPreparedSourceBinding The immutable workspace-task snapshot consumed by one Eval client. EvalTrialSummary Reconstructible aggregate counts for fixed outer Eval trials. FrontendCoRendering The fixed co-rendering requirement shared by compatible frontend plans. GatewayFrontendBinding The stable schema branch for a Gateway-only frontend binding. GatewayPdRouterFrontendBinding The stable schema branch for a fused Gateway and P/D Router binding. GatewayPlan The public frontend component plan returned independently from Engine and
P/D Router planning. HttpActionSpec An HTTP action invoked against a logical serving endpoint. HttpTargetRegistryReadiness The integration-owned HTTP registry contract for target-aware readiness. IntegrationIdentity The integration’s identity recorded on its results: adapter id, adapter
version, and the framework it lowers to. LaunchFileDeclaration One immutable text input a rendered process requires before it can launch. MeasurementDataAssetCacheStore One effective physical cache store as reported by its owning runtime. MeasurementDataAssetContentEntry One member of a complete, ordered content closure. MeasurementDataAssetPreparationRequest One narrowly scoped source-preparation request issued before measurement
materialization. The selected measurement definition remains authoritative
for dataset semantics. MeasurementDataAssetPreparationResult Terminal result returned by the owning measurement runtime. MeasurementDataAssetVerification Expected-versus-observed integrity evidence for qualified source content. MeasurementDataFiles A normalized mapping from dataset split names to selected file patterns. MeasurementModelInput The model identity used by measurement clients. Unlike integration
planning, a benchmark client may need a controller-visible tokenizer
locator. MeasurementProtocol The schema root aggregating the request and result surfaces used by the
product-owned Eval and Bench measurement clients. Its optional fields are
code-generation anchors and are never all populated in one message. Parallelism Framework-neutral component-aware parallelism ([[RFC-0003:C-SERVE-PARALLELISM]]).
Every component is optional so an operator can override one without
repeating the rest; omitted components are filled by the integration into a
complete effective shape. ParallelismAttention Attention-block parallelism: tensor, data, and context degrees. ParallelismExperts MoE expert parallelism: tensor, data, expert, and dense-tensor degrees. ParallelismOuter Outer deployment parallelism: tensor and pipeline degrees. PdRouterPlan The P/D orchestration component plan, kept separate even when the same
process and provider also realize Gateway. PdRoutingPolicies Independent policies for choosing prefill and decode targets. PlanServeInput The requested serve shape a PlanServe operation lowers into a topology. PlanServeResult The lowered topology returned by a PlanServe: effective Engine roles,
whole-replica requirements, logical links, and separate optional Gateway
and P/D Router component plans. ProcessSpec A launchable process: its argument vector and environment. RawArtifact A raw output file a client produced, retained as workload evidence. RenderInputDeclaration One workspace-authored UTF-8 source file an integration declares during
planning for the control plane to supply during final rendering. RenderServeInput The planned topology plus the control plane’s concrete allocations that a
RenderServe operation turns into final process invocations. RenderServeResult The final process invocations returned by a RenderServe, one per supplied
allocation. ServeModelInput The logical model supplied during serving planning and rendering. ServeReplicaRequirement A whole-replica resource and readiness requirement the integration declares
without choosing placement, ranks, or concrete endpoints. ServeRoleInput A requested serving role: its identity, kind, replica cardinality, and
declared (not-yet-completed) parallelism and settings. ServeRoleResult A role as the integration resolved it: preserved identity and cardinality
with the complete effective settings and parallelism. ServerMetricsEndpointInput The resolved server-metrics endpoint supplied to a measurement client. ServerMetricsEndpointRequirement An integration-owned logical server-metrics endpoint. SuppliedRenderInput The original declared path plus the exact UTF-8 contents and digest the
control plane supplies to an integration during final rendering. SyntheticAcceptanceCurveInput The curve form’s wire payload ([[RFC-0003:C-SERVE-SYNTHETIC-ACCEPTANCE]]). SyntheticAcceptanceOutcome The integration’s resolved synthetic acceptance outcome, returned in the
plan_serve response whenever the request carried the declaration
([[RFC-0006:C-INTEGRATIONS]]). AdapterErrorCode Machine-readable failure category an adapter reports in an AdapterError AdapterRequest The one JSON request an integration reads from stdin, tagged by the
requested operation. AdapterResponse The one JSON response an integration writes to stdout, tagged by outcome. AdapterResult The successful result of an AdapterRequest AllocationLaunch The machine-local launch channel selected by the control plane. AuxiliaryModelKind The governed auxiliary weight artifact kinds a serve declaration may attach
([[RFC-0003:C-SERVE-AUXILIARY-MODELS]]). BenchAgenticAcquisitionOutcome BenchArtifactLevelInput The artifact level a Bench case requests from the measurement runtime.
Omission resolves to diagnostic, which retains the full raw
request/response export; performance keeps only normalized per-request
records and the summary export ([[RFC-0004:C-BENCH-ARTIFACT-LEVEL]]). BenchCacheStartInput BenchDatasetCacheState Read-only cache state observed while resolving the Bench plan. BenchLoadInput How a Bench case paces its requests. BenchPrefixSharingInput One declared exact final-prompt prefix geometry. BenchPromptInput The frozen prompt representation, route, and rendering authority for a
serving Bench population. BenchPromptRouteInput The named OpenAI-compatible endpoint family derived from prompt authority. BenchPromptTemplateSource Where the concrete template used for local prompt-length projection came
from. The model server remains authoritative for transport rendering. BenchRenderingAuthorityInput The owner that turns semantic prompt input into the final prompt tokens. BenchRequestRepresentationInput The request payload representation frozen by prompt resolution. BenchRequestSourceInput One closed request origin lowered by Inferlab for the Bench runtime. BenchSharedSystemContentInput One pre-template shared system-content declaration for server-rendered
synthetic chat. BenchTokenDistributionKindInput BenchTokenSelectorInput One fixed or bounded inclusive-uniform token-length selector. CaptureMechanism The workload-attached capture mechanism selected for a profiled serving
topology ([[RFC-0004:C-WORKLOAD-PROFILING]]). CaptureWindowControlEndpoint The logical workload endpoint exposing a capture target’s window-control
actions. ClientStatus The terminal outcome a measurement client reports. EndpointProtocol The application protocol a workload endpoint speaks. EvalDefinitionInput The measurement an Eval client runs against the workload endpoint. EvalFailureKind A typed Eval failure category preserved across the client boundary. EvalMetricComparison The threshold comparison selected from lm-eval’s scoring direction. EvalMetricGateConclusion The terminal conclusion of an lm-eval metric threshold gate. EvalPromptInput The prompt rendering authority for a generative lm-eval definition. EvalTaskSourceInput The resolved lm-eval task source consumed by the release-owned Eval runner. FrontendComponents The only two frontend bindings the adapter protocol accepts. Tuple representation
deliberately serializes as the closed ordered arrays ["gateway"] and
["gateway", "pd_router"] rather than as an open component list. FrontendGatewayComponent The literal first member of every closed frontend component binding. FrontendHandoff The currently demonstrated Gateway-to-P/D Router handoff is internal to
one fused frontend process. FrontendPdRouterComponent The literal second member of the fused P/D frontend component binding. FrontendProcessRole The one canonical process role available to a frontend. GatewayTarget The target a Gateway forwards accepted public requests to. HttpMethod The HTTP method of an action specification. KvTransferMechanism The KV-transfer mechanism connecting prefill and decode. MeasurementDataAssetAcquiredSource The complete immutable source identity established by preparation. MeasurementDataAssetCacheOutcome Read-only or preparation-time outcome for one independently owned cache. MeasurementDataAssetEffectiveSelection The effective selection reported by the owning measurement runtime. MeasurementDataAssetPreparationNextPhase The next separately durable preparation phase selected by the owning
runtime after source resolution. MeasurementDataAssetPreparationPhase One externally observable source-preparation phase. Separating resolution
from acquisition lets the control plane durably commit each result. MeasurementDataAssetReadiness Whether preparation established a closed source or an explicitly opaque
non-reproducible source. MeasurementDataAssetRemoteMetadataOutcome Whether preparation contacted a remote metadata authority. MeasurementDataAssetSourceBytesOutcome Whether immutable source bytes were reused or downloaded. MeasurementDataAssetSourceInput The measurement-owned source whose bytes or task selection are prepared. PromptCacheReadZeroRepresentation How a backend with cache reporting enabled represents a zero-token cache read. ProtocolVersion The shared protocol version used by framework integrations and release-owned
measurement clients. The only accepted value is 10 (serialized as the
string "10"); a mismatch is rejected before lowering. ReadinessProbe How the control plane decides a process is ready. RenderSource Which bounded implementation renders an accepted concrete frontend
allocation. This is a lowering boundary only; the control plane always
owns placement, lifecycle, cleanup, endpoints, and records. RenderedServeProcess A rendered process bound to the model-rank or frontend allocation identity
it was produced for. ServeProcessAllocation One concrete process allocation supplied to RenderServe. Model-rank and
process-only frontend identities are distinct so a frontend cannot acquire
model coordinates or a model locator by construction. ServeRoleKind The logical role a replica plays in a serve topology. ServeRoleLink A directed link between serve roles the integration declares as part of the
topology. ServeTopology The serving deployment topology. SettingValue A framework-specific server setting value carried as structured JSON data
(never a pre-rendered shell fragment) across the integration boundary. SyntheticAcceptanceInput The synthetic acceptance declaration carried to the selected integration
([[RFC-0003:C-SERVE-SYNTHETIC-ACCEPTANCE]]): exactly one of the explicit
acceptance length or the digest-verified golden curve with its lookup
coordinates. The draft count is not declared; the integration determines
it from the operator’s speculative configuration ([[ADR-0043]]). TargetEndpointScheme The application protocol used to identify serving targets in a registry. MEASUREMENT_SCHEMA_ID MEASUREMENT_WIRE_SOURCE PROTOCOL_SCHEMA_ID PROTOCOL_WIRE_SOURCE measurement_schema protocol_schema