VORQ DocsCapacity
How many jobs a provider may hold, how the network grants it, and how per-model limits shape what the daemon offers.
Requested and granted capacity
provider.capacity is the number of jobs the daemon asks to hold at once. It sends the request
at startup. Left unset, it is the sum of what each model entry may hold (see below).
The network grants:
granted = max(1, min(requested, ceiling) × reputation / 1000)Onboarding sets your ceiling and your starting reputation when it registers you. Reputation ranges from 100 to 1000. It rises by 5 for every settled job and falls by 40 for every job failed after the 300-second grace window or reclaimed after a missed SLA. A provider registered at 1000 fills its whole capacity from its first claim. One registered at 200 and asking for 4 slots holds 1 slot until about 60 settled jobs, 2 until about 110, and all 4 at full reputation.
The daemon reads the grant every sweep and never offers more than it. When the grant is below
capacity, it logs the network grants N of the M slots configured, and
vorqd_capacity_granted shows the grant.
What each poll offers
For each model, the poll's free is the smallest of:
- the daemon's slots (capacity or grant, whichever is smaller) minus jobs held;
- what the model entry's own limits admit.
Without a rate_limit window at the entry's shortest SLA, the entry admits what it could start
now: concurrency minus jobs held, and the slack in its tightest rate_limit window. With such
a window, the entry may hold that window's budget, less the attempts already started in it and
one reserved for every job still waiting. Shorter windows then only pace the starts. This is
how a 24h entry can take a day's work in the morning and run it concurrency at a time.
A tripped breaker makes free zero for that model. Jobs recovered after a restart are resumed
regardless of free capacity: they were already sold.
Sizing
A missed SLA costs reputation, and reputation costs capacity. Set capacity and limits below
what the backend sustains inside the SLA window, not at its peak. Use
vorqd_backend_latency_seconds and vorqd_job_duration_seconds to measure real throughput and
leave headroom. Settling fewer jobs cleanly beats claiming more and missing some.