Docs · Task 06 · Docs for 1.0.12

Let another machine do the work

A laptop with no model of its own can ask a machine that has one, over your own encrypted link, between devices you have already paired. The asking machine gets an answer; the serving machine does the inference.

Step 1

Pair the two machines first

This runs over the same link as everything else in your workspace, so the machines have to know each other. Add a second device covers the pairing and the short code that makes it safe.

Step 2

Turn sharing on, on the machine that will serve

On the machine with the model, open Settings → Privacy & data → Compute sharing and choose Allow members. The same panel offers Stop sharing, and it is the only switch involved; there is no standing grant to forget about.

Compute sharing panel: mode Allow members, Stop sharing offered, jobs run for
                others this session 9, requests approved this session 9, and a list of recent
                jobs each showing the capability, the peer's generated name and the time.
The serving machine's view. Both session counters read 9, and the list below shows each job with the peer it ran for and when. Captured on v1.0.11.
Step 3

Ask from the other machine

Nothing changes on the asking side. Ask as you normally would; if that machine has no model of its own, the work goes to the paired machine that does and the answer comes back. The serving machine's counters go up and the job appears in its recent list.

A machine works out whether a model will actually fit in the memory it has free, not in the memory it has in total. A laptop that cannot really run one no longer tries to download it when a paired machine can answer instead.

✓
Measured, not asserted. In the session above, one machine served 9 jobs for 2 paired devices: two real machines over the mesh, not two profiles on one computer.
Step 4

What this does not do yet

We have run this between two machines. With more than two paired and able to serve, we are not yet confident the right one answers every time, so treat a larger set as early rather than finished. The recent-jobs list also labels every row authorised, including rows that ran to completion; if your two counters ever disagree, that list will not tell you which rows did the work.

This is your own machines helping each other. A person you work with can join the same workspace with a key and ask questions against it: they get answers, never your files.

What crosses, and what is refused

LeavesThe question, as an embedding, and the passage that answers it. Not the document. The machine holding the file reads it locally and returns only the part that answered.
RefusesA peer you have not admitted is refused, visibly from both sides. The refusal deliberately does not enumerate what exists, so an unknown service name and an ungranted peer look identical — a stranger learns nothing by asking.
CostsNothing. Each serve writes a signed record naming the node that did the work, so "a peer answered" is never the whole of what you can check.

One run, two halves

When part of a job is answerable here and part needs the tier you pointed it at, Armore runs both and then tells you which half ran where. Two halves that ran in the same place is not a split, however it was planned, and is not reported as one.

If a half came back from a host it could not verify, the run reads unverifiable — not local. Armore will not tell you your data stayed on your device unless it can prove it did, because saying so is a positive claim and an unprovable dispatch has not earned it.

When a half is refused, whose limit was it?

Every refusal names an actor. These are the exact words the card shows you.

ReasonWhat the card saysWhose
cloud-not-provisionedNo provider was connectedOperator
cloud-grade-unprovenThe provider never proved it was good enoughOperator
cloud-allowance-exhaustedYour cloud allowance is spentYou
cloud-spend-cap-reachedWe hit our own spending ceilingOurs
consent-absentYou had not allowed this job to leave the deviceYou
consent-revoked-mid-jobYou withdrew permission while this was runningYou
timed-outThe cloud half ran out of timeYou
dispatch-failedThat half never startedNobody
empty-local-generationThe on-device model returned nothingNobody

Two of those blame nobody, deliberately. dispatch-failed adds "It failed before anything was sent. Nothing left the device for it" — a claim about your data staying put that it is entitled to make, because the half never left. empty-local-generation adds "Not a refusal and not an answer": the model on your machine produced no text, which is a third thing and is reported as one.