Skip to content
Docs

Control Plane & Dashboard

Everything Lifeboat does is available in a browser. The point is operational: private inference should not require a person who knows the right command-line invocations to be awake, and day-to-day operation should not require SSH access to GPU hosts.

What does the dashboard show?

Section titled “What does the dashboard show?”
Area What it covers
Dashboard System summary, GPU fleet visualization with live utilization, load-balancer pool status, and recent activity
Models The registry — add, download with progress, inspect, choose GGUF variants, see quantization at a glance, delete
Servers Inference-server list and detail — create, start, stop, force-stop, test, read logs, bulk actions
Cluster Fleet nodes, named clusters, and a fleet-wide GPU table
Deploy Multi-node deployment of a model across a named cluster
API Keys Issue, pause, resume, revoke, and delete keys for inference traffic
Tokens The encrypted Hugging Face credential vault for gated and private models
Users Accounts and role assignment
Audit A paginated, filterable log of every administrative action

Is there a way to test a model without writing code?

Section titled “Is there a way to test a model without writing code?”

Yes. A built-in playground sends requests to a running server from the browser, with streaming output, so an operator can confirm a newly deployed model behaves correctly — and see reasoning output and errors — before any application is pointed at it.

Can it be operated without a CLI?

Section titled “Can it be operated without a CLI?”

Yes — the browser control plane is the intended operating surface, and no CLI is required for normal work. A command-line entry point exists for non-container installations and for scripted bring-up, and every control-plane action is also available through the REST API for automation. See API and automation.

What monitoring does Lifeboat expose?

Section titled “What monitoring does Lifeboat expose?”
  • Prometheus metrics at a standard endpoint, covering dispatch counts, rejections by reason, in-flight requests, queue depth, and queue wait-time distribution — scrapeable by an existing Prometheus or any compatible agent.
  • Live GPU telemetry per node, refreshed continuously and shown in the fleet view.
  • Load-balancer state as a live snapshot, including actual-versus-configured traffic share per server.
  • Per-server logs readable in the UI, with a mountable log directory for longer retention.
  • Audit log of every administrative action, filterable by action, target, time range, and user, and covering sign-ins, permission denials, and all configuration changes.

Each model process additionally exposes its own metrics on its own port for teams that want engine-level detail.

Sources and references

Section titled “Sources and references”