Skip to content
ignisignis

Everything in one binary

A serving engine is only as useful as the things around it: a way to try a prompt, a way to see what the card is doing, a way to fetch the weights, a way to change a setting. In ignis those are in the same ignis-server binary, started with the same command. There is no second service to deploy for the UI, no exporter to attach for metrics, and no script to download the model. Every setting is one declaration that is a flag, an environment variable and a file key at once.

How it works

Section titled “How it works”

The Playground. A React app embedded into the server at build time and served at /ui/: chat with tools and subagents, parallel sessions, image input, the Image studio, the Decide tab, the Monitor and the Server view. Its sessions are stored in the browser (IndexedDB), not on the server (ADR 0026).

Configuration. Every option is declared once (ADR 0046) and is a flag --<group>-<field>, an IGNIS_<GROUP>_<FIELD> variable and a key <group>.<field> in YAML or JSON, with per-model values and a built-in profile for the card. GET /v1/config shows the running configuration, and PATCH /v1/config changes it, reloading the model only for a field that needs a reload. The reference tables are generated from the code, and cargo test fails if they drift from it.

Counting. POST /v1/tokenize counts a chat body’s prompt as it would be served, with its template, tools and thinking controls, without serving it.

Metrics. Prometheus has its own loopback listener, keyless and on by default, and the Monitor reads the same set.

Fetching. Started with no artifact, a GPU build asks, downloads and verifies the model (ADR 0033).

Access. server.api_key takes a key or auto, which generates one and prints it once. server.expose opens a public quick-tunnel URL and always requires a key (ADR 0028). The /v1 surface also documents itself at /v1/openapi.json (ADR 0036).

Platforms. A v* tag publishes a Windows .zip, a Linux .tar.gz and a linux/amd64 container image on ghcr.io, all compiled for SM120a.

The Monitor, live, from the same binary

The numbers

Section titled “The numbers”

There are no performance figures here. These are the defaults:

Piece Default Reference
Playground on, at /ui/ server.ui
Metrics listener on, 127.0.0.1:9464, no key server.metrics_bind
API key unset: the API is open on localhost server.api_key
Public tunnel off; when on, a key is always required server.expose
Model download a verified catalog entry, or a mirror download.endpoint

See it, control it

Section titled “See it, control it”
  • Switch pieces off: --server-ui false, --server-metrics false.
  • Print every field: ignis-server help --fields; GET /v1/config for the running values.
  • Routes: /ui/, /v1/config, /v1/tokenize and /v1/openapi.json.
  • CLI: ignis-server model download <id>, and ignis-bench for measurements.
  • Metrics: the full list is in the metrics, and Telemetry and metrics shows how to scrape and read them.
  • Playground: the Monitor draws the metrics live and the Server view edits the configuration.

Limits

Section titled “Limits”
  • A binary built without the Playground’s web/dist serves /ui/ as a page that says how to build it (The Playground).
  • The Prometheus listener is keyless and on loopback. Exposing it is the operator’s act.
  • A quick tunnel is a quick tunnel: the key is required, and GET /v1/config never shows it.
  • ignis targets SM120a, the RTX 5090 and the RTX PRO 6000. It does not run on a DGX Spark (DGX Spark).

Read more

Section titled “Read more”