The advanced path

For an operator who wants to know what each knob trades, and set it. Every setting is one field with a flag, an env var and a config-file key; the explanation pages say what a mechanism does and why, the configuration pages which field moves it.

Understand the engine

Start with Why ignis: each topic there says what was measured, on what, and where it stops, and links the page below that explains its mechanism.

1. How ignis serves a request — lanes, prefill and decode, the KV pool, speculation, one model at a time.
2. The VRAM plan — how a load lays out every byte before it starts, and why a start is refused rather than run out of memory.
3. The KV cache and prompt reuse — pages, the pool, prefix reuse and retained state, the host tier and KV-disk.
4. Speculative decoding — DFlash2 for the 27B, MTP for Flash-Next, and when each pays.
5. Models and artifacts — the 27B, Flash-Next and the image model, and what each needs in VRAM and host RAM.

Configure it

6. Configuration — fields and groups, precedence, per-model values, the config file, profiles, live changes.
7. Profiles and defaults and the built-in profiles — what each card’s profile chooses.
8. The fields by topic: context and speculation, Flash-Next MTP, VRAM and the KV pool, KV-disk, the n-gram cache, host RAM, vision, prompting.
9. Every config field — the generated list, with each field’s default, rule and whether it changes live.

Run it for real

10. Access — keys, CORS origins, the metrics listener, public exposure.
11. Telemetry and metrics and the metric list.
12. Air-gapped machines, an operator catalog, Knative.
13. Model switching and idle unload — what a switch costs, and when the server frees the card.

Tip

ignis-server config print shows what a start would run — file, env vars and flags resolved — and the Server view of the Playground shows where each running value came from.

Caution

A raised context or a lowered headroom is checked by the VRAM plan at load: a value that does not fit refuses the start and names what it needed. Change one thing at a time and read the refusal (Troubleshooting).