Docs

The workbench is a thin client over a local open-lease server. It never holds your provider credentials; it just calls the server’s REST API and streams chat through its OpenAI-compatible proxy.

You’re running the bundled workbench, so it already talks to the server that served it. One command boots the API, the proxy, and this UI together:

gpu ui

To point the browser build at a remote server instead, host it and run gpu serve --cors-origin https://<this-page-origin> on the machine with the GPUs.

Every deployment is a card: its state, the GPU, uptime, the cost accrued so far (ticking live between polls), and the endpoint once it’s serving. The list refreshes about once a second, so you watch a deployment climb to Ready in real time. Click a card for the full detail view: the event log, the state timeline, health, and logs.

Two ways in. Catalog model picks a validated entry and its recommended GPU. Any HF model deploys any vLLM-servable Hugging Face repo ad hoc (the CLI equivalent is gpu deploy --hf-repo); you choose the GPU and, optionally, a context length.

The form shows the hourly rate and whether a data center has capacity right now. Deploy is non-blocking: a background daemon drives the pod to Ready. If none is running, start one with gpu up.

Pick a Ready deployment and chat with it. Messages stream token-by-token through the OpenAI-compatible proxy, which routes by model name to the right pod. Enter sends; Shift+Enter adds a newline; Stop aborts an in-flight response.

A deployment moves through these on its way to Ready and back down:

• RequestedQueued. The reconciler will acquire a pod on the next tick.
• ProvisioningRenting the GPU pod from the provider.
• BootingThe pod is powering on.
• DownloadingPulling the model weights (the slow step on a cold start).
• StartingvLLM is loading the model into GPU memory.
• ReadyServing. Use it in the Playground or via the API.
• DegradedWas ready, now failing health checks. Still billing.
• StoppingTearing the pod down.
• StoppedTorn down. No longer billing.
• FailedGave up after retries. Check the event log on the detail page.

A pulsing dot means it’s working toward Ready. A pod bills from Provisioning through Degraded (anytime it exists on the provider), not once it’s Stopped or Failed.

Open any deployment’s detail page for an API panel: the base URL, the model name, and copy-paste cURL / Python / JavaScript that POST a chat completion. It’s OpenAI-compatible, so any OpenAI SDK works by pointing its base_url at your server’s /v1 and using the deployment’s model name.

• Your data stays local. The server URL and token live in this browser’s storage. Requests go straight to your machine, never through a third party.
• The token guards everything. If you set --api-token, every management and inference route requires it.
• Cross-origin is opt-in. A hosted workbench can only reach your server if you start it with --cors-origin for that exact page. It’s off by default and never wildcarded, so a running server isn’t exposed to other sites.