Deciding when your GPU works

The question that decides whether someone leaves a GPU connected is not "how much does it pay" — it is "what is it doing when I want to use my computer, and what does it do to my electricity bill". Both are settings, and you can change them from either end.

The two places

$ mahout config show                          # what it is set to, and where that came from
$ mahout config keys                          # every setting, with an explanation
$ mahout config set pause_on_fullscreen=false
$ mahout config set max_power_watts=250 disk_quota_bytes=500GB

Or in the dashboard, on the machine's own row. Same document, same values, either direction.

What you can decide

Availability windows

Empty means always available, which is the default. Give it a schedule and the machine takes work only inside those hours:

$ mahout config set windows='[{"cron":"0 22 * * 1-5","duration_minutes":540}]'

That reads: start at 22:00 on weekdays, run for nine hours. In the dashboard the same thing is one line — 0 22 * * 1-5 for 540m — and it is deliberately still a cron expression rather than a day-picker, so both sides describe the schedule identically. A widget that produced a slightly different meaning from the same clicks would be worse than a little syntax.

Overnight is the obvious use: cheaper electricity on many tariffs, and nobody wants the fans up while they are in the room.

A power cap

max_power_watts holds the card below its own limit. Lower is cooler, quieter and cheaper to run — and slower, and since you are paid per token produced, slower means less. A genuine trade, which is why it is yours to make rather than ours. Leave it empty for the card's default.

VRAM you keep

vram_reserve_bytes is held back from the engine for everything else you run on that card. If you keep a browser with forty tabs and a video editor open, reserving a couple of gigabytes is the difference between a smooth desktop and a stuttering one.

Which models, and how much disk

models_allowed restricts the machine to specific model ids; empty — the usual answer — means everything the card is qualified for. disk_quota_bytes is the cache budget for weights; a larger cache means fewer re-downloads when work shifts between models.

Pausing for you

pause_on_user_activity and pause_on_fullscreen are both on by default: work stops within seconds of a keypress, and while any game or fullscreen application is running. You do not have to remember. See keeping your machine yours for the rest of what is under your control.

Why both sides can write

Because both are real. You are at the machine sometimes and on your phone other times, and a system where one of those is read-only is a system where you end up SSH-ing into your own desktop to turn something off.

Two writers need a rule, and ours is that nobody's change disappears without being told. Every save says which version of the settings it was editing. If the other side changed something first, the save is refused and you are shown what the settings actually are now, so you can make your change again on top. The alternative — last write wins — quietly loses an edit whenever the agent syncs a stale copy a second after you pressed save, and you would never know.

Both the CLI and the dashboard show where the last accepted change came from, so "did that take effect" is answerable at a glance.

One honest caveat: windows are stored, synced and displayed today, and the agent's evaluation of them lands in the next release. Everything else on this page is enforced now.