Central servers · Docs

What the central server does

A job has two parts: the GPU workers you pick, which do the heavy work, and one central server that coordinates them. You never pick or pay for the central server separately. The platform provides one with every job.

Job type	Central server	What it does
Inference	Ray head	Holds the endpoint and checks the API key on every request, queues requests, forwards each one to a worker and returns the answer. Runs no model.
Training	FL server	Runs the federated rounds: after each round it collects the weights every worker trained, merges them, and sends the merged weights back to each worker. Reports the job status. Your data never leaves the workers.

Inference   caller ──► central server ──► worker ──► answer to caller
Training    workers train ──► central server merges ──► workers get new weights

How a central server is chosen

When you launch, the platform picks a central server for the job automatically:

A server reserved for you comes first, if you have one that fits the job.
Otherwise a shared server with enough free memory and disk is used.
Under Filters in the submit wizards you can set minimum memory and disk for it, or ask for a dedicated server that runs only this job.

If no server meets the job's minimums, the launch stops with a message saying so. The wizard then offers to order one.

Defaults and failures

	Inference	Training
Default minimum	2 GB free RAM, 10 GB free disk	1 GB free RAM, no disk minimum
Raise it for	Large uploaded weights (disk), busy endpoints and long streams (dedicated)	Full-weight models above ~1B: ~12 bytes of RAM per parameter, dedicated
A worker drops out	Its replicas move to the other workers	The round can’t finish and the job fails
The central server goes down	The endpoint stops; resubmit the job	The job fails; resubmit it

Sizing details: Choosing hardware for inference and training.

Getting one reserved for you

Order a central server from Support → Request a central server, from My orders, or straight from the workers step of either submit wizard. Say whether it is for inference or training and, optionally, how much memory, disk and CPU it needs.

Support prepares the server and reserves it for you. The request page shows its status and the conversation.
Once reserved, only your jobs run on it, and they use it without any selection at launch.
My orders lists your reserved servers and which of your jobs are running on each.
Cancel the request when you no longer need it; the server is then shared again.

note

Reserved central servers are not billed today. You pay for the GPU workers your jobs use.

PreviousPort a centralized recipe to FL