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 aggregating every wire type. It exists to generate one
committed JSON schema (and the Python SDK models); its optional client
fields are never all populated in a single message. 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. BenchDefinitionInput The workload shape a Bench client drives, shared across its load cases. CaptureControlRequirement The HTTP paths a capture target exposes to open and close a capture window. CaptureTargetRequirement Marks a replica as a profiling capture target and carries its window
control ([[RFC-0004:C-WORKLOAD-PROFILING]]). 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. EndpointRequirement The workload endpoint’s protocol and API path, plus an optional
prefix-cache-reset action a Bench case can invoke between runs. 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. HttpActionSpec An HTTP action (method and path) invoked against the workload endpoint. IntegrationIdentity The integration’s identity recorded on its results: adapter id, adapter
version, and the framework it lowers to. 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. PlanServeInput The requested serve shape a PlanServe operation lowers into a topology. PlanServeResult The lowered topology returned by a PlanServe: the effective server-level
shape, per-role resolution, whole-replica requirements, role links, and the
public and per-role endpoint requirements the control plane then allocates. ProcessSpec A launchable process: its argument vector and environment. RawArtifact A raw output file a client produced, retained as workload evidence. 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. RenderedServeProcess A rendered process bound to the allocation id it was produced for. ServeModelInput The model to serve: its resolved locator and the name it is served under. ServeProcessAllocation One concrete process the control plane placed, supplied to RenderServe:
its identity, rank, machine, devices, model locator, and allocated
endpoints and named ports. 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. 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 BenchLoadInput How a Bench case paces its requests. 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. HttpMethod The HTTP method of an HttpActionSpec KvTransferMechanism The KV-transfer mechanism connecting prefill and decode. ProtocolVersion The adapter protocol version an integration speaks. The only accepted value
is 3 (serialized as the string "3"); a mismatch is rejected before
lowering. PublicEndpointRequirement How the topology’s public workload endpoint is exposed. ReadinessProbe How the control plane decides a process is ready. 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. PROTOCOL_SCHEMA_ID protocol_schema