Glossary
Alphabetical reference of all Wingsmith terms and abbreviations.
This page lists terms used across Wingsmith’s code and docs. For how they fit together, see the concepts overview and architecture.
A
Admission. The phase a deployment passes through before it is placed: pending → admitted → placed → running. For serving, admission succeeds immediately unless quota refuses it. See Deployments.
Adoption. Recording a workload the operator found already running in a cluster, without taking control of it. See Deployments.
Agent (hangar-agent, libs/agent). The per-cluster process that dials out to the gateway and carries inference requests to model frontends inside the cluster. Not the operator. See Clusters.
Allocation. Observed: the devices currently committed to a replica. See Devices.
Artifact. Immutable, content-addressed model weights (a base model, a fine-tune, a quantized derivative or a LoRA adapter), with one or more storage locations. See Artifacts and Workloads.
Authority. A field on a deployment: observed (recorded and checked for drift, never written to the cluster) or managed (the plane may change it). See Deployments.
C
Cache salt. A per-tenant value the gateway adds to each text request (nvext.cache_salt) so Dynamo’s KV/prefix cache is never shared across tenants. See Tenants and Projects.
CacheEntry. Observed: whether an artifact’s weights are resident on a cluster or node.
Cluster. A Kubernetes cluster enrolled in Wingsmith. Declared as a registry entry, with observed state (reachability, sync state, operator version) reported by its operator. See Clusters.
Component. Something installed in a cluster and versioned independently: the operator, the agent, the serving runtime, exporters.
Console (libs/console). The web UI. A client of the plane’s HTTP API, like hgr.
Control plane / plane (libs/plane). The central service: the HTTP API, the hub and the reactor, over one Postgres database. Stores intent; never in the inference request path. See How It Works.
D
DCGM / DCGM exporter. NVIDIA’s GPU telemetry exporter. Wingsmith reads device identity (GPU UUIDs) and health from it. See Devices.
Declared / observed / operation. The three kinds of object. Declared objects are written by the plane and obeyed by the cluster; observed objects are written by the cluster and read by the plane; operations are imperative, long-running and can fail. See Operations.
Deployment. One workload on one cluster, carrying that cluster’s share of the scale. The unit the operator reconciles. Not a Kubernetes Deployment. See Deployments.
Device. One accelerator, identified by its UUID, with health, memory, MIG profile and current allocation. See Devices.
Device class. Wingsmith’s name for a kind of accelerator (for example an 80 GB H100 SXM). Vendor strings map to device classes through libs/packages/contract/devices.json. See Devices.
DGD (DynamoGraphDeployment). NVIDIA Dynamo’s custom resource describing a serving graph: frontend, workers and their roles. The operator renders each Wingsmith deployment into one.
Disaggregated serving. Running prefill and decode as separate worker roles, each with its own replica count and devices. A workload declares its roles; an aggregated workload has one.
Dynamo (NVIDIA Dynamo). The inference serving framework Wingsmith renders workloads into. It runs vLLM, SGLang or TensorRT-LLM as the engine and provides the frontend, router and KV-aware routing.
E
Endpoint. The stable public model name clients call. Plane-only; resolves to one or more deployments with weights or a failover order. The gateway reads endpoints, never workloads. See Endpoints.
Engine. The inference runtime inside a worker: vLLM, SGLang or TensorRT-LLM.
Enrolment token. The per-cluster credential minted when a cluster is registered. The operator presents it to the hub and the agent presents it to the gateway. It identifies exactly one cluster. See Clusters.
EP. Expert parallelism: how mixture-of-experts models split experts across devices. See TP / PP / EP.
F
Frontend. Dynamo’s OpenAI-compatible HTTP entry point for one model’s deployment, which routes to that deployment’s workers. Wingsmith renders one per model.
G
Gateway (libs/gateway). The one Wingsmith process in the inference request path. Takes OpenAI-compatible requests, picks a cluster from the routing table, and forwards over the tunnel. See How It Works.
H
hgr (libs/cli). The command line client for the plane’s API. See hgr CLI Reference.
hgrbench (libs/benchmarking). Runs several load generators against an OpenAI-compatible endpoint and reports them side by side. See hgrbench.
hgrconform (libs/conformance). Probes an OpenAI-compatible endpoint and compares its behaviour with a reference provider. See hgrconform.
hgrtransit (libs/transit). Measures the gateway, tunnel and agent path on one host, with no GPU or cluster. See hgrtransit.
Hub. The plane process that terminates the operators’ WebSocket streams, with resume and backpressure. See How It Works.
I
Internal / public listener. The gateway’s two client-facing listeners. The internal one has no authentication and must not be reachable from outside your network; the public one identifies the tenant from the Host header and can require an API key. See Tenants and Projects.
K
KV cache / prefix cache. Attention state kept by the engine so repeated prompt prefixes are not recomputed. Configured on the workload; unrelated to the weight cache. See Artifacts and Workloads.
M
Metrics label. The value of the cluster label a cluster’s metrics carry in the central metrics store, declared on the cluster’s registry entry. See Clusters.
MIG (Multi-Instance GPU). NVIDIA’s partitioning of one GPU into isolated slices. See Devices.
O
Operator (hangar-operator, libs/operator). The per-cluster process that receives desired state from the hub, stores it in CRDs, renders it into Kubernetes and Dynamo objects, and reports observed state back. Also a principal kind in the plane’s auth. See Clusters.
Operation. Imperative, long-running, failable action: drain, scale, cordon, reboot, device reset, cache warm, rolling update, component upgrade. See Operations.
P
Placement. The plane’s choice of clusters for a workload and each cluster’s share of its scale. Picks clusters, never devices. See Quotas and Placement.
Policy. Placement constraints (regions, residency, device classes, allowed clusters) attached to a tenant, endpoint, workload or cluster. See Quotas and Placement.
Pool. A declared provisioning spec for a homogeneous set of nodes in a cluster.
PP. Pipeline parallelism: how model layers split across devices. See TP / PP / EP.
PriorityClass. A named priority for workloads. Carries a preemptible flag that is recorded but not yet acted on. See Quotas and Placement.
Project. A unit inside a tenant that device quota attaches to. See Tenants and Projects.
Q
Quota. A ceiling on how many devices of one class a project may hold, either in one cluster or across the fleet; a cluster’s own quota takes precedence over a fleet-wide one there. Checked by placement at admission. See Quotas and Placement.
R
Reactor. The plane process (or loop inside the hub) that derives alerts and reconciles derived configuration, under a lease.
Replica. Observed: one running instance of a deployment on named devices. See Deployments.
remote_write. The Prometheus protocol each cluster’s metrics agent uses to push metrics to the central store.
S
Serving configuration. The engine, version, quantization, parallelism, context length, cache settings and engine arguments of a workload. Hashed into its serving_digest. See Artifacts and Workloads.
ServingQuota. A declared traffic commitment for one endpoint (request rate, token rate, concurrency). Recorded and shown, not yet enforced.
StreamCheckpoint. The CRD in which the operator persists its resume cursor for the hub stream.
SyncedResource. The CRD in which the operator stores the desired state it received from the hub, so a cluster keeps its intent across operator restarts and plane outages.
T
Tenant. An organization using the installation’s models. Owns endpoints, API keys and the host name its requests arrive on. Also a principal kind in the plane’s auth. See Tenants and Projects.
TopologyDomain. Observed: which devices share an NVLink/NVSwitch fabric, PCIe root or NUMA node. See Devices.
TP / PP / EP. Tensor, pipeline and expert parallelism: how a model is split across devices. TP splits within layers, PP splits layers across devices, EP splits experts (for mixture-of-experts models).
Tunnel (libs/packages/tunnel). The HTTP/2 connection the agent dials to the gateway, over which the gateway sends requests into the cluster. See How It Works.
W
Workload. What you asked to run, globally: an artifact, its serving configuration, scale bounds, priority and placement constraints. Has one deployment per cluster it runs on. See Artifacts and Workloads.
Code names vs UI names
Wingsmith is the public product name. Code identifiers stay as hangar:
hgr: CLI toolhangar-operator: Operator image and processhangar-agent: Agent image and processHANGAR_*: Environment variables (e.g.,HANGAR_GATEWAY_SERVICE_TOKEN)hangar.dev: CRD API group (interim, subject to change before release)
All CRDs, config keys, chart names, image names, and API paths use hangar names. Only UI and prose use Wingsmith.