Skip to main content

Pick a device

Your device is registered. This part finds it in the fleet, reads what the platform knows about it, and walks the device page, which is where you will come back whenever a run behaves oddly.

If your team already had devices before you arrived, everything here works the same. Pick any one that is online, rather than the one you added in the previous part.

Read the fleet

Open Devices. The list is the organization's whole fleet, shared by every project.

The Devices list, with status, name, hardware, operating system and architecture, tags and last-seen columns

Each row carries what you need to choose one: whether it is online, what hardware it has (cores, memory, GPU and VRAM where there is one), its operating system and architecture, and when it was last seen. Pick one that is online and not busy. A workstation-class Linux machine finishes a first run fastest; a phone or a small board works too, and takes longer.

StateWhat it means for your benchmark
Online or IdleReady to take work.
BusyRunning a job already. Yours will queue behind it.
OfflineNo heartbeat for 60 seconds. It cannot take work.
SuspendedTaken out of service by an operator.
DirtyA previous job's teardown failed. New benchmark jobs are blocked until it is cleared.

Read the device page

Select the device you picked.

A device page showing status, tags, device info and live health

The header carries the status and how long ago the device was seen, with Terminal and Remote Desktop beside it. Device Info is the hardware inventory the agent detected when it registered. Health is live, refreshed by a heartbeat every 15 seconds: CPU, memory and disk use, temperature, uptime, and history over the last day.

The tabs are the rest of what you can do with the machine.

  • Overview is the page above: inventory, live health, health history, and the device's tags.
  • Jobs is everything the platform has run on this device, newest first, each with its state and its result. After the next two parts, your benchmark appears here.
  • Services is what the device keeps running: a process per row with its state, uptime and restart count, and a live log you can open per service. This is where a model deployment shows up on the device that serves it.
  • Files is a file manager for the device. Browse its filesystem, upload from your machine or straight from the artifact library, download a file, and save what you downloaded back into the library as a new artifact. Large uploads are chunked and resume after a dropped connection.
  • Commands runs a one-off command and records it with its exit code, its output and the time it ran, so a command you ran last week is still readable this week.

Two buttons in the header do the interactive work. Terminal opens a shell in the browser, a real PTY on Linux, macOS and rooted Android. Remote Desktop opens the device's screen where a provider exists for that platform; when one cannot be started, the device says why and names the remedy rather than hiding the button.

All of it travels over the single connection the agent opened outward to the platform. Nothing connects inward to the device, which is why a machine behind a home router or a corporate firewall works with no ports opened.

What you have now

A device the platform can reach, and enough of its page to recognize a healthy one. The next part registers the model you want to measure.