Configuration

Every key of vorqd.yaml — the provider block, model entries, backend limits and load-time validation.

vorqd.yaml has two top-level keys: provider and models. Any string value may contain env:NAME, which is replaced by the environment variable NAME at load. A referenced variable that is not set stops the daemon with an error naming it.

provider:
  wallet_key: env:VORQ_WALLET_KEY
  box_key: env:VORQ_BOX_KEY
  api_url: https://api.vorq.co
  capacity: 4

models:
  - model: deepseek-ai/deepseek-v4-pro:fp8
    modality: text
    slas:
      "24h": { rate_in: "160000", rate_out: "550000" }
    backend:
      preset: openai-chat
      base_url: http://localhost:8000/v1
      model: deepseek-ai/DeepSeek-V4-Pro

provider

Key	Default	Meaning
`api_url`	`https://api.vorq.co`	Base URL of the VORQ coordinator.
`wallet_key`	`$VORQ_WALLET_KEY`	Operator wallet private key (secp256k1, hex). Signs the session and every operation. The daemon exits if no key is available.
`box_key`	required; forbidden when confidential	Curve25519 private key (64 hex characters) that opens designated payloads. Its public half must match your provider record. Must be absent when any model sets `confidential: true`.
`capacity`	derived	Jobs the daemon requests to hold at once; a positive integer. Unset, it is the sum over models of each entry's `rate_limit` window at its shortest SLA, else its `concurrency`; an entry with neither makes it required.
`metrics_port`	`9090`	Port of the ops server (`/healthz`, `/metrics`), bound on all interfaces.
`log_level`	`info`	`debug`, `info`, `warn` or `error`.
`poll_interval_s`	`5`	Seconds between poll sweeps.
`safety_margin_s`	`60`	Seconds before the SLA deadline after which no attempt runs and no settle is tried.
`fail_grace_s`	`300`	Window after a claim in which a `fail` is penalty-free, mirroring the job registry. Used for logging only.
`max_input_bytes_per_unit`	`16`	For `text` and `embedding` bids with a nonzero `rate_in`: most payload bytes one declared input unit may buy (plus 4096 bytes of slack). Larger bids are not claimed. `0` disables the check.
`state_db`	`vorqd-state.sqlite`	SQLite file for the handles of jobs in flight at resumable backends. Relative paths resolve against the working directory. Must be writable.
`escrow_url`	`api_url`	Base URL of the escrow that releases payload keys for open bids.
`pricing`	off	Private acceptance floor. See pricing.
`bid_filter.min_age_s`	none	Accepted for compatibility and ignored, with a warning at startup.

There is no provider id key: the daemon learns its id at sign-in.

pricing

Key	Default	Meaning
`max_discount_pct`	`0`	Discount at load 0, shrinking linearly to 0 at load 1. Integer, 0–90.
`bid_tolerance_pct`	`0`	Extra discount while load is below `low_load_pct`. Integer, 0–90.
`low_load_pct`	`30`	Load threshold in percent, 0–100.

The total discount is capped at 90%. It applies only to models with a load block. Unknown keys are refused.

Model entry

Each item of models:

Key	Default	Meaning
`model`	required	Network model name clients submit against. Must be in the coordinator's catalog.
`slas`	required	Map of SLA window to `{rate_in?, rate_out}`. See slas.
`backend`	required	How to run a job: exactly one of `preset` (presets) or `request` (raw mappings), plus the shared backend keys.
`modality`	none	`text`, `embedding`, `image` or `video`. Used when the catalog names none. The daemon claims nothing for a model whose modality is unknown.
`confidential`	`false`	`true` makes the daemon a confidential provider. See Run a confidential provider.
`load`	none	Load source for dynamic pricing. See load. A sibling of `backend`; a `load` key inside `backend` is refused.
`sla_backends`	none	Map of SLA window to a backend patch. See sla_backends.

slas

Keys are windows: a whole number and a unit s, m, h or d ("1h", "24h").
rate_out is required on every window; rate_in is optional.
Rates are whole numbers (strings or integers) of atomic token units per 106 units of work, at most 2128−1. A decimal is refused.
One ask is published per window. The rates are also the claim floor for that window.
At most 64 asks (windows across all models) fit in one ask book.

load

Either a Prometheus probe:

Key	Default	Meaning
`url`	required	Absolute `http(s)` URL of a Prometheus text endpoint.
`metric`	required	Series name. `:` and `_` are treated as equal. The busiest sample wins.
`scale`	`1.0`	Positive divisor that maps the series to 0..1.

or {source: occupancy}, which reads the entry's own fullness and takes no other key. It needs concurrency or rate_limit on the backend.

The probe is scraped once per sweep with a 2-second timeout. A reading older than 30 seconds, a failed scrape, or a series with no finite non-negative sample counts as unknown, which earns no discount.

sla_backends

Each value is merged key by key over backend and validated as a backend of its own. Patch keys win; null removes a key; preset in a patch drops an inherited request/response, and request drops an inherited preset. Windows must exist in slas. Limit keys are refused in a patch. The model's asks are published only while every one of its backends passes its health check.

Backend keys

These keys are valid on every backend block, preset or raw.

health

Key	Meaning
`health.path`	Probe URL, requested with `GET` once per sweep. A path starting with `http` is used as is; otherwise it is appended to `base_url` (presets) or to the scheme and host of `request.url` (raw mappings). A status below 400 is healthy. While unhealthy, the model's asks are withdrawn.

Backend limits

Key	Default	Meaning
`retries`	`0`	Attempts after the first. A `429`, any `5xx`, a transport error or timeout, and statuses in `retry_statuses` are retried; any other `4xx` fails the job. Non-negative integer.
`retry_statuses`	none	Extra `4xx` codes to retry.
`retry_backoff_s`	`30`	First retry delay, doubling after each retry. A `429`'s `Retry-After` (in seconds) wins when longer. Positive number.
`retry_backoff_max_s`	`900`	Ceiling on one delay. Must be at least `retry_backoff_s`.
`timeout_s`	remaining SLA budget	Read timeout of one attempt, never beyond `deadline − safety_margin_s`. With `stream: true`, the longest silence between chunks.
`concurrency`	unbounded	Most requests in flight for this model. Positive integer.
`rate_limit`	none	Map of window to the most attempts started in it, retries included. The window equal to the shortest SLA is also how many jobs the entry may hold. The loader refuses a hold budget that the shorter windows cannot start inside the SLA.
`trip_after`	`3`	Consecutive jobs failed with the backend at fault before the model's asks are withdrawn. `0` disables.
`trip_cooldown_s`	`60`	Seconds the model stays withdrawn after tripping. Positive number.

No wait or attempt is scheduled past deadline − safety_margin_s; the job is failed back instead. queue is not a key and is refused.

Media policy

param_caps, reference, resolutions, durations, auto_duration and adaptive_aspect apply to image and video models on any backend shape. See Raw mappings: media policy.

Load-time validation

The daemon validates the whole file before it contacts anything and exits with status 2 and a message naming the key and model on the first violation. Beyond the types and ranges above:

models is a non-empty list; each backend holds exactly one of preset or request.
A raw response.result holds exactly one of text, media_urls or media_b64, and a text result also maps completion_tokens.
A raw poll holds exactly one of status_url or request; handle requires request, and a poll request with handle cannot use {submit.…}. A poll request using {handle} needs handle.
max_polls is a positive integer.
stream is a boolean and true only on openai-chat.
headers on a preset is a string map and is refused on openai-batch.
service_tier is flex or priority; completion_window is a non-empty string; endpoint starts with /.
param_map, reasoning, param_caps and params_supported follow the shapes in Presets.
prepare steps, response.ok, poll.failure_code and poll.failure_message follow the shapes in Raw mappings.
reference, resolutions, durations, auto_duration and adaptive_aspect follow the shapes in media policy.

Further checks run as the daemon starts, and stop it on failure: preset must be one of openai-chat, openai-embeddings, openai-responses or openai-batch; every model must be in the coordinator's catalog; and a regular provider's box key must match its provider record.

Examples

vllm.yaml: a self-hosted OpenAI-compatible text model.
queue-backend.yaml: an image model on a submit-then-poll API.
confidential.yaml: a confidential provider.

providerCopy Anchor Link

pricingCopy Anchor Link

Model entryCopy Anchor Link

slasCopy Anchor Link

loadCopy Anchor Link

sla_backendsCopy Anchor Link

Backend keysCopy Anchor Link

healthCopy Anchor Link

Backend limitsCopy Anchor Link

Media policyCopy Anchor Link

Load-time validationCopy Anchor Link

ExamplesCopy Anchor Link

On this page