Give important applications priority
Give higher-priority applications precedence when requests are waiting for capacity. Rank your applications once, and the most important keep getting through longest when you're busy.
Before you begin
You need at least two applications, one per workload: for example a coding agent and a nightly batch job. The priority list appears once you have two.
1. Give each workload its own application
In Applications, choose Add application, name it for the workload, and choose Create key. Use that key only for that workload. A new application starts at the bottom of the list.
2. Set the order
In Applications, under Application priority, drag the most important application to the top. Or open an application's menu and choose Move up or Move down. With a keyboard, focus a row's handle, press Space, move it with the arrow keys, and press Space again.
From the terminal, list your keys with their ids and weights, then set a weight. A higher weight ranks higher. Keep weights from 100, where a new application starts, to 1000, as the list does. Below 100, that application is turned away sooner when you're busy:
saylek settings account keys list
saylek settings account keys reweight KEY_ID --weight 800
You know it worked when the list shows your applications in the order you set.
What the order does
Give higher-priority applications precedence when requests are waiting for capacity. That is what the order sets. When you're busy, lower-ranked applications are turned away first and higher-ranked ones keep getting through longest. It applies only to your account's own requests. For example, with Coding agents above Nightly QA, a waiting coding-agent request goes ahead of a waiting QA request. A request waits for a machine up to 180 seconds when it streams and up to 60 seconds when it does not, or less if your application gives it less time.
It does not:
- stop a request that is already running;
- reserve capacity, or make a request faster;
- rank your requests against anyone else's.
It is also not the rule that lets a machine's owner go first on their own machine, which Make more of one GPU explains.
If something goes wrong
| What you see | What to do |
|---|---|
| No Application priority list | You have one application. Add a second |
reweight refuses the weight | Use a whole number up to 1000 |
| A lower-ranked application is refused (HTTP 429) while you're busy | If the error's details name consumer, that is the order working. Move it up, or spread its requests out. key_deprioritized means its weight is below 100: set it to 100 or more |
A request ends with an error that names wait_elapsed | No machine was free in time. Retry later, or add a machine |
| Important requests still wait | Priority orders your waiting requests; it adds no capacity. Add a machine, as Add machines and GPUs describes |
Reference
- Connect your app: keys and client setup.
- Set up Saylek for a team: who sets priority when several people share machines.
- CLI reference: every command.
Last checked 2026-09-25 · read as markdown at /docs/prioritize-applications.md