Settings

Platform switches, training images, the hardware registry, and scheduling windows.

Administrators only. Everything here is backed by a deployment setting — the page is a view onto the configuration reference, not a second source of truth.

Platform settings, grouped by concern.

Sections

Section	What it holds
Platform switches	Master switches that change behaviour platform-wide
Sessions and tokens	Lifetimes for every kind of token. Shortening one expires credentials already issued
Limits and caching	Upload size, fetch caching, config refresh interval
Training runtime	Default images per framework, and automatic retry
Inference and deployments	Playground and managed deployment settings
Hardware	The GPU series and profile registry
Schedule	Per-series time windows

Unsaved changes are counted in the header, so a half-finished edit is visible rather than lost.

Hardware registry

This is what makes --profile h200:8 mean something.

Field	What it defines
Series id	The name users type: h200, h100
Display name	What the console shows
Nodes / GPUs per node	The shape a profile requests
Weight	Relative cost, used in fair-share accounting
Process env	KEY=VALUE lines injected into jobs on this series

Changing a series or profile that currently has active jobs is confirmed separately. Those jobs already hold an allocation described by the old definition; the platform will not silently reinterpret it underneath them.

Scheduling windows

Per-series rules saying which weekdays and hours that hardware may be used. A window can cross midnight.

Setting	Effect
Enforcement off	Windows are displayed but nothing is refused
Minimum remaining time	A job is not dispatched into a window with less than this left
Stop grace	How long running jobs get after a window closes

Validation is strict at save time — a rule with no weekday selected, or a window missing an end time, is rejected with the series named. A half-defined window that silently never matches would be worse than no window at all.

Maintenance mode

Turning it on queues new submissions without dispatching them, and drained jobs stay paused with auto-resume set. Turning it off resumes automatically: each paused job is resubmitted under its original run id and continues from its checkpoint.

The note you write is shown to anyone whose submission is held. Write one — “maintenance” with no end time in sight generates support questions that the note would have answered.