Every config field

Note

Generated from the code by make docs-reference (crates/server/src/user_docs.rs); an edit here is overwritten, and cargo test fails while this page and the code disagree.

Every setting is one field with the same name on every surface: the flag --<group>-<field>, the env var IGNIS_<GROUP>_<FIELD> and the config-file key <group>.<field>. Precedence, highest first: the per-model flag, the flag, the per-model env var, the env var, the per-model file key, the file key, the profile, the default. Configuration explains the model; ignis-server help --fields prints this same table from the binary you run.

server

server.bind

Flag	--server-bind
Env var	IGNIS_SERVER_BIND
Kind	text
Default	127.0.0.1:8000
Models	all
/v1/config	shown by GET, not patchable

The API’s listen address (host:port). Localhost on the OpenAI port by default; never changed at runtime — a rebind is a restart.

server.api_key

Flag	--server-api-key
Env var	IGNIS_SERVER_API_KEY
Kind	key|auto
Default	unset
Models	all
/v1/config	hidden from GET, not patchable

The key every /v1 request must present as Authorization: Bearer <key>; auto generates one at start and prints it (the only time a key is printed). Unset keeps the API open, as it has always been on localhost. Never shown by GET /v1/config.

server.expose

Flag	--server-expose
Env var	IGNIS_SERVER_EXPOSE
Kind	expose mode
Default	unset
Rule	one of cloudflare-quick
Models	all
/v1/config	shown by GET, not patchable

Expose the server beyond its bind address: cloudflare-quick opens a public https://*.trycloudflare.com URL, printed at start. An exposed server always requires an API key — auto when none is set.

server.metrics

Flag	--server-metrics (may stand bare: true)
Env var	IGNIS_SERVER_METRICS
Kind	bool
Default	true
Models	all
/v1/config	shown by GET, not patchable

Serve Prometheus metrics on their own listener (metrics_bind), and at /ui/metrics unless the Playground is off. On by default; false serves neither.

server.metrics_bind

Flag	--server-metrics-bind
Env var	IGNIS_SERVER_METRICS_BIND
Kind	text
Default	127.0.0.1:9464
Models	all
/v1/config	shown by GET, not patchable

The metrics listener’s address: loopback by default, no API key, never exposed, never the API’s bind; refused with metrics off. The default address taken (a second server on the host) is a warning, and the server starts without the listener; another address that cannot be bound refuses the start.

server.ui

Flag	--server-ui (may stand bare: true)
Env var	IGNIS_SERVER_UI
Kind	bool
Default	true
Models	all
/v1/config	shown by GET, not patchable

Serve the Playground at /ui/. A binary built without web/dist serves a page saying how to build it.

server.request_timeout

Flag	--server-request-timeout
Env var	IGNIS_SERVER_REQUEST_TIMEOUT
Kind	seconds
Default	2400
Rule	0..=3600
Models	all
/v1/config	shown by GET, patchable live

Seconds a request answered whole may run, from its admission (an image job’s from its start), before its deadline ends it as timeout_reply says: chat completions, the Responses API over HTTP, each question of a decision, an image job, and a request’s media acquisition. 0 is no deadline.

server.stream_timeout

Flag	--server-stream-timeout
Env var	IGNIS_SERVER_STREAM_TIMEOUT
Kind	seconds
Default	2400
Rule	0..=3600
Models	all
/v1/config	shown by GET, patchable live

Seconds a request that streams may run, from its admission (an image job’s from its start), before its deadline ends it as timeout_reply says: chat completions and the Responses API streamed, a Responses WebSocket’s response, a streamed image job. 0 is no deadline.

server.timeout_reply

Flag	--server-timeout-reply
Env var	IGNIS_SERVER_TIMEOUT_REPLY
Kind	reply
Default	partial
Rule	one of partial, error
Models	all
/v1/config	shown by GET, patchable live

What a request past its deadline is answered: partial, what was generated by then as an answer a length limit cut (finish_reason: "length", or a response incomplete with max_output_tokens); error, a 504 request_timeout or a stream’s error ending. A decision’s question and an image job are an error either way.

server.socket_idle_close

Flag	--server-socket-idle-close
Env var	IGNIS_SERVER_SOCKET_IDLE_CLOSE
Kind	seconds
Default	300
Rule	0..=604800
Models	all
/v1/config	shown by GET, patchable live

Seconds a Responses WebSocket may sit idle — no frame from the client and no response running or queued on it — before the server closes it with a normal close frame saying why; the client’s next request opens another. An open socket holds a Knative pod, and its GPU, from scaling down. 0 never closes.

server.system_message_policy

Flag	--server-system-message-policy
Env var	IGNIS_SERVER_SYSTEM_MESSAGE_POLICY
Kind	policy
Default	merge
Rule	one of merge, strict
Models	all
/v1/config	shown by GET, patchable live

Where system messages go: merge joins a leading run into the system prompt and keeps a later one as its own block in place; strict answers 400 for a system message that is not first.

server.developer_message_policy

Flag	--server-developer-message-policy
Env var	IGNIS_SERVER_DEVELOPER_MESSAGE_POLICY
Kind	policy
Default	inplace
Rule	one of inplace, into-system, after-system, one-after-system, reject
Models	all
/v1/config	shown by GET, patchable live

Where developer messages go: inplace, into-system, after-system, one-after-system or reject. A leading developer message is the system prompt except under reject.

model

model.id

Flag	--model-id
Env var	IGNIS_MODEL_ID
Kind	text
Default	unset
Models	all
/v1/config	shown by GET, not patchable

The id the model is served under. Unset, a load is served under its own model’s id (qwen3.8-27b or qwen3.8-flash-next); without an artifact, qwen3.8-27b. An id naming the other model than the artifact’s is refused.

model.artifact

Flag	--model-artifact
Env var	IGNIS_MODEL_ARTIFACT
Kind	path
Default	unset
Models	all
/v1/config	shown by GET, not patchable

The .ninfer artifact to load. Unset, the model is looked for under download.path, and fetched when it is not there.

model.max_context

Flag	--model-max-context
Env var	IGNIS_MODEL_MAX_CONTEXT
Kind	tokens
Default	40960
Rule	>= 1
Models	all
Per model	--qwen38-model-max-context, --qwen38flashnext-model-max-context
/v1/config	shown by GET, patchable, with a model reload

The largest prompt + generation a single request may reserve, in tokens. On the 27B, bounded by its attention’s envelope for the KV format.

model.default_max_tokens

Flag	--model-default-max-tokens
Env var	IGNIS_MODEL_DEFAULT_MAX_TOKENS
Kind	tokens
Default	38912
Models	all
Per model	--qwen38-model-default-max-tokens, --qwen38flashnext-model-default-max-tokens
/v1/config	shown by GET, patchable, with a model reload

The max_tokens of a request that sends none, its reasoning included, never past what its prompt leaves of max_context; an explicit cap always wins. 0 = none, up to the context.

model.kv_format

Flag	--model-kv-format
Env var	IGNIS_MODEL_KV_FORMAT
Kind	kv format
Default	hq-e8-2b
Rule	one of bf16, hq-e8-2b, hq_e8_2b, hq
Models	all
Per model	--qwen38-model-kv-format, --qwen38flashnext-model-kv-format
/v1/config	shown by GET, patchable, with a model reload

The KV storage format: hq-e8-2b (the serving default) or bf16 (the retained oracle format).

model.prefill_attention

Flag	--model-prefill-attention
Env var	IGNIS_MODEL_PREFILL_ATTENTION
Kind	prefill attention
Default	fp8
Rule	one of fp8, fp16
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

The 27B’s prompt attention precision (GitHub #335): fp8 (the default, the band pass in 8 bits) or fp16 (the FP16-accumulate fallback). Flash-Next’s attention is QSA and has no such pass.

model.exception_weights

Flag	--model-exception-weights
Env var	IGNIS_MODEL_EXCEPTION_WEIGHTS
Kind	exception weights
Default	bf16
Rule	one of bf16, fp8
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

How the 27B holds its nine BF16 exception tensors on the device (GitHub #341): bf16 (the default, as the artifact stores them) or fp8, converted at load to FP8 row-scale for the FP8 tensor cores. Flash-Next has none.

model.prefill_chunk

Flag	--model-prefill-chunk
Env var	IGNIS_MODEL_PREFILL_CHUNK
Kind	tokens
Default	1024
Rule	nonzero multiple of 128
Models	all
Per model	--qwen38-model-prefill-chunk, --qwen38flashnext-model-prefill-chunk
/v1/config	shown by GET, patchable, with a model reload

The prefill chunk width, in tokens.

model.decode_share

Flag	--model-decode-share
Env var	IGNIS_MODEL_DECODE_SHARE
Kind	percent
Default	unset
Rule	0..=99
Models	all
Per model	--qwen38-model-decode-share, --qwen38flashnext-model-decode-share
/v1/config	shown by GET, patchable, with a model reload

The percent of the model’s time decoding lanes keep while a prompt prefills. Unset, the model’s own: 25 on both. 0 is one decode round per chunk, 50 splits time evenly.

model.enable_thinking

Flag	--model-enable-thinking (may stand bare: true)
Env var	IGNIS_MODEL_ENABLE_THINKING
Kind	bool
Default	true
Models	all
/v1/config	shown by GET, patchable live

Whether a request that says nothing thinks.

model.reasoning_effort

Flag	--model-reasoning-effort
Env var	IGNIS_MODEL_REASONING_EFFORT
Kind	reasoning effort
Default	unset
Rule	one of none, minimal, low, medium, high, xhigh, max
Models	all
/v1/config	shown by GET, patchable live

The reasoning_effort of a request that names none. Unset, the template’s own default.

model.thinking_budget

Flag	--model-thinking-budget
Env var	IGNIS_MODEL_THINKING_BUDGET
Kind	tokens|off
Default	32768
Models	all
/v1/config	shown by GET, patchable live

The reasoning tokens a request may spend before the model’s close is forced; off = no budget. A request’s own thinking_budget overrides it.

model.loop_guard

Flag	--model-loop-guard
Env var	IGNIS_MODEL_LOOP_GUARD
Kind	loop guard mode
Default	rewind
Rule	one of off, log, rewind
Models	all
/v1/config	shown by GET, patchable live

The loop guard (ADR 0052): rewind, a loop in a request’s reasoning (chat completions, /v1/responses) is taken back and the model nudged, then closed after loop_guard_retries nudges; log, it is only logged; off, no guard, as benches and gates run for reference parity.

model.loop_guard_ratio

Flag	--model-loop-guard-ratio
Env var	IGNIS_MODEL_LOOP_GUARD_RATIO
Kind	permille
Default	265
Rule	1..=999
Models	all
/v1/config	shown by GET, patchable live

A window of scored reasoning whose zlib ratio is under this many thousandths is a loop.

model.loop_guard_window

Flag	--model-loop-guard-window
Env var	IGNIS_MODEL_LOOP_GUARD_WINDOW
Kind	chars
Default	1000
Rule	500..=100000
Models	all
/v1/config	shown by GET, patchable live

The loop guard’s window, in characters of scored reasoning.

model.loop_guard_normalize

Flag	--model-loop-guard-normalize (may stand bare: true)
Env var	IGNIS_MODEL_LOOP_GUARD_NORMALIZE
Kind	bool
Default	true
Models	all
/v1/config	shown by GET, patchable live

Whether the loop guard scores the reasoning normalized: fenced and inline code taken out, every run of digits read as one 0. off scores it as written, which at this window and ratio flags far more than the raw prototype did (3,000 and 120).

model.loop_guard_raw_gate

Flag	--model-loop-guard-raw-gate
Env var	IGNIS_MODEL_LOOP_GUARD_RAW_GATE
Kind	permille
Default	300
Rule	0..=999
Models	all
/v1/config	shown by GET, patchable live

The loop guard’s raw gate, in thousandths: a loop’s last loop_guard_window characters as written must score under it too, which a table of numbers (a template once its digits fold) does not. 0, no gate.

model.loop_guard_clean

Flag	--model-loop-guard-clean
Env var	IGNIS_MODEL_LOOP_GUARD_CLEAN
Kind	permille
Default	400
Rule	1..=999
Models	all
/v1/config	shown by GET, patchable live

A window of scored reasoning at or over this many thousandths is clean: a rewind goes back to the start of the last one before the loop (no clean one since the last nudge: to its end).

model.loop_guard_margin

Flag	--model-loop-guard-margin
Env var	IGNIS_MODEL_LOOP_GUARD_MARGIN
Kind	chars
Default	0
Rule	0..=100000
Models	all
/v1/config	shown by GET, patchable live

How many characters before that clean window’s start a rewind cuts.

model.loop_guard_retries

Flag	--model-loop-guard-retries
Env var	IGNIS_MODEL_LOOP_GUARD_RETRIES
Kind	retries
Default	2
Rule	0..=10
Models	all
/v1/config	shown by GET, patchable live

The nudges a looping reasoning gets before the guard closes it.

model.loop_guard_persist

Flag	--model-loop-guard-persist
Env var	IGNIS_MODEL_LOOP_GUARD_PERSIST
Kind	checks
Default	4
Rule	1..=100
Models	all
/v1/config	shown by GET, patchable live

How many checks in a row must score a loop before the guard acts: a loop persists, a list or a log transcribed into the reasoning ends with its content. 1, at the first.

model.loop_guard_min_kept

Flag	--model-loop-guard-min-kept
Env var	IGNIS_MODEL_LOOP_GUARD_MIN_KEPT
Kind	chars
Default	1
Rule	0..=100000
Models	all
/v1/config	shown by GET, patchable live

A rewind that would keep fewer characters of reasoning than this is only logged, and that run of loop scores is not acted on: a reasoning that opens in a loop, or in an enumeration scored as one, would otherwise be dropped whole. 0, always rewind.

model.loop_guard_holdback

Flag	--model-loop-guard-holdback
Env var	IGNIS_MODEL_LOOP_GUARD_HOLDBACK
Kind	chars
Default	0
Rule	0..=100000
Models	all
/v1/config	shown by GET, patchable live

In rewind mode, the characters of reasoning held back from the client: a rewind whose cut falls in them reaches it with no tag and none of the loop. 0, none; under ~2,000 it hides almost no rewind.

model.loop_guard_check_every

Flag	--model-loop-guard-check-every
Env var	IGNIS_MODEL_LOOP_GUARD_CHECK_EVERY
Kind	chars
Default	250
Rule	1..=100000
Models	all
/v1/config	shown by GET, patchable live

The loop guard scores the window again after this many characters of new reasoning.

model.loop_guard_snap

Flag	--model-loop-guard-snap
Env var	IGNIS_MODEL_LOOP_GUARD_SNAP
Kind	chars
Default	300
Rule	0..=100000
Models	all
/v1/config	shown by GET, patchable live

A rewind’s cut moves back to the line break before it when one is within this many characters. 0, never.

model.loop_guard_min_continuation

Flag	--model-loop-guard-min-continuation
Env var	IGNIS_MODEL_LOOP_GUARD_MIN_CONTINUATION
Kind	tokens
Default	512
Rule	>= 1
Models	all
/v1/config	shown by GET, patchable live

The least of a request’s max_tokens a continuation must have left after the cut and the nudge: with less, the loop is only logged.

model.loop_guard_submit_wait

Flag	--model-loop-guard-submit-wait
Env var	IGNIS_MODEL_LOOP_GUARD_SUBMIT_WAIT
Kind	seconds
Default	60
Rule	0..=3600
Models	all
/v1/config	shown by GET, patchable live

How long a continuation the engine refuses as full waits for room, in seconds, before the loop is left as it is; 0 tries once. A WebSocket request’s continuation waits in the socket’s admission queue instead, as the socket’s own requests do.

model.loop_guard_nudge

Flag	--model-loop-guard-nudge
Env var	IGNIS_MODEL_LOOP_GUARD_NUDGE
Kind	text
Default	Wait, I'm starting to repeat myself. Let me stop and move on with what I already have.
Models	all
/v1/config	shown by GET, patchable live

The sentence, in the model’s own voice, the loop guard’s first rewind puts after the kept reasoning, on a paragraph of its own.

model.loop_guard_nudge_again

Flag	--model-loop-guard-nudge-again
Env var	IGNIS_MODEL_LOOP_GUARD_NUDGE_AGAIN
Kind	text
Default	I keep repeating myself. I already have enough to write the answer, so I'll wrap up my reasoning now.
Models	all
/v1/config	shown by GET, patchable live

The sentence every later rewind puts there, until loop_guard_retries nudges close the reasoning.

model.rope_scaling

Flag	--model-rope-scaling
Env var	IGNIS_MODEL_ROPE_SCALING
Kind	rope scaling
Default	auto
Models	all
Per model	--qwen38-model-rope-scaling, --qwen38flashnext-model-rope-scaling
/v1/config	shown by GET, patchable, with a model reload

The text rotary table: auto (the default), none (the linear table) or yarn:F[,t=..][,bf=..][,bs=..], which rescales the checkpoint’s trained 262,144-position envelope — what a context past it needs. auto decides per model from its model.max_context: none up to 262,144, else YaRN at max_context / 262,144. Both models run the table.

model.idle_unload

Flag	--model-idle-unload
Env var	IGNIS_MODEL_IDLE_UNLOAD
Kind	seconds
Default	300
Rule	0..=604800
Models	all
/v1/config	shown by GET, patchable live

Seconds with no request using the model after which it is unloaded — VRAM, pinned host memory and KV-RAM given back — while the server stays up; the next request loads it again and waits for it. 0 never unloads.

model.host_margin_bytes

Flag	--model-host-margin-bytes
Env var	IGNIS_MODEL_HOST_MARGIN_BYTES
Kind	bytes
Default	4G
Models	all
Per model	--qwen38-model-host-margin-bytes, --qwen38flashnext-model-host-margin-bytes
/v1/config	shown by GET, patchable, with a model reload

What a text model’s host plan must leave of the physical memory available at load (free plus standby): a load whose pinned lines — the 27B’s token embedding, Flash-Next’s experts and n-gram hot rows, staging, retained host slots, the KV-RAM arena — leave less is refused, naming the line to shrink, so the desktop is never paged out. The image model makes no host plan: its start carries it for a switch to a text model.

model.adapt_context_step

Flag	--model-adapt-context-step
Env var	IGNIS_MODEL_ADAPT_CONTEXT_STEP
Kind	tokens
Default	65536
Rule	nonzero multiple of 64
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

The 27B’s adaptation (ADR 0063): when its VRAM plan refuses a max_context the profile or the default gave, the load lowers it this many tokens at a time, never below this many – a start whose plan does not hold one step is refused. A whole number of KV pages (64 tokens).

vram

vram.kv_pool_bytes

Flag	--vram-kv-pool-bytes
Env var	IGNIS_VRAM_KV_POOL_BYTES
Kind	bytes|tokens
Default	unset
Models	all
Per model	--qwen38-vram-kv-pool-bytes, --qwen38flashnext-vram-kv-pool-bytes
/v1/config	shown by GET, patchable, with a model reload

The KV pool: a byte count (K/M/G suffix read) or tokens as <n>tok, <n>Ktok, <n>Mtok. Unset, the KV pool policy’s: the rest of the VRAM budget when every weight is on the device (the 27B); 524,288 tokens shared by the lanes when Flash-Next’s experts stream, the rest to the expert cache. Refused below one max_context sequence and a page per retained slot.

vram.headroom_bytes

Flag	--vram-headroom-bytes
Env var	IGNIS_VRAM_HEADROOM_BYTES
Kind	bytes
Default	1536M
Models	all
Per model	--qwen38-vram-headroom-bytes, --qwen38flashnext-vram-headroom-bytes
/v1/config	shown by GET, patchable, with a model reload

What a derived VRAM budget leaves to the desktop and every other process on the card: the budget is the memory free at start less this. Not with budget_bytes.

vram.budget_bytes

Flag	--vram-budget-bytes
Env var	IGNIS_VRAM_BUDGET_BYTES
Kind	bytes
Default	unset
Rule	>= 1
Models	all
Per model	--qwen38-vram-budget-bytes, --qwen38flashnext-vram-budget-bytes
/v1/config	shown by GET, patchable, with a model reload

The device memory the whole process may hold, weights included; refused above free memory. Unset, derived from headroom_bytes.

vram.allow_oversubscription

Flag	--vram-allow-oversubscription (may stand bare: true)
Env var	IGNIS_VRAM_ALLOW_OVERSUBSCRIPTION
Kind	bool
Default	false
Models	all
Per model	--qwen38-vram-allow-oversubscription, --qwen38flashnext-vram-allow-oversubscription
/v1/config	shown by GET, patchable, with a model reload

Start above free memory, with a warning. Needs budget_bytes.

vram.allow_expert_cache_below_floor

Flag	--vram-allow-expert-cache-below-floor (may stand bare: true)
Env var	IGNIS_VRAM_ALLOW_EXPERT_CACHE_BELOW_FLOOR
Kind	bool
Default	false
Models	Qwen3.8-Flash-Next
/v1/config	shown by GET, patchable, with a model reload

Start Flash-Next with an expert cache below its 12 GiB floor, with a warning — decode slows sharply below it.

vram.pinned_expert_bytes

Flag	--vram-pinned-expert-bytes
Env var	IGNIS_VRAM_PINNED_EXPERT_BYTES
Kind	bytes
Default	0
Models	Qwen3.8-Flash-Next
/v1/config	shown by GET, patchable, with a model reload

Flash-Next’s calibration-hottest experts held in the expert cache for the whole load, outside its LRU, with no copy in the pinned host pool: whole experts, hottest first, until they reach this many bytes. The pool takes that much less RAM; the LRU has those slots fewer (on the 5090: nothing measurable at 1 GiB, ~1% of three-lane decode at 2 GiB). 0 pins none; the rtx5090 profile pins 2 GiB. A card that holds every expert keeps none in RAM anyway and ignores it.

vram.adapt_spare_tokens

Flag	--vram-adapt-spare-tokens
Env var	IGNIS_VRAM_ADAPT_SPARE_TOKENS
Kind	tokens
Default	65536
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

The 27B’s adaptation (ADR 0063): at more than one decode lane, a context the load lowers is kept only when the KV pool holds it and a spare reservation of this many tokens, room for a second request; 0, one sequence is enough. At one lane the plan’s own floor decides.

reuse

reuse.prompt

Flag	--reuse-prompt (may stand bare: true)
Env var	IGNIS_REUSE_PROMPT
Kind	bool
Default	true
Models	all
Per model	--qwen38-reuse-prompt, --qwen38flashnext-reuse-prompt
/v1/config	shown by GET, patchable, with a model reload

Cross-request state reuse. Off, no prompt checkpoint is captured or claimed, and no prefix is shared unless retained_device or retained_host gives slots for it.

reuse.retained_device

Flag	--reuse-retained-device
Env var	IGNIS_REUSE_RETAINED_DEVICE
Kind	slots
Default	unset
Models	all
Per model	--qwen38-reuse-retained-device, --qwen38flashnext-reuse-retained-device
/v1/config	shown by GET, patchable, with a model reload

Retained slots in VRAM, reserved in the VRAM plan and handed out first. Unset, 0. With prompt reuse off, a count shares heads between live siblings only.

reuse.retained_host

Flag	--reuse-retained-host
Env var	IGNIS_REUSE_RETAINED_HOST
Kind	slots
Default	unset
Models	all
Per model	--qwen38-reuse-retained-host, --qwen38flashnext-reuse-retained-host
/v1/config	shown by GET, patchable, with a model reload

Retained slots in one pinned host block reserved at start, ~222 MiB each, a PCIe copy per capture and per claim. Unset, the model’s own: two per decode lane on the 27B, 8 on Flash-Next; 0 with prompt reuse off.

reuse.retained_interactive_ttl

Flag	--reuse-retained-interactive-ttl
Env var	IGNIS_REUSE_RETAINED_INTERACTIVE_TTL
Kind	seconds
Default	300
Models	all
Per model	--qwen38-reuse-retained-interactive-ttl, --qwen38flashnext-reuse-retained-interactive-ttl
/v1/config	shown by GET, patchable, with a model reload

Idle seconds after which a main-conversation checkpoint in KV-RAM ranks as a subagent’s. Needs prompt reuse on.

reuse.kv_host_pool_bytes

Flag	--reuse-kv-host-pool-bytes
Env var	IGNIS_REUSE_KV_HOST_POOL_BYTES
Kind	bytes
Default	2G
Models	all
Per model	--qwen38-reuse-kv-host-pool-bytes, --qwen38flashnext-reuse-kv-host-pool-bytes
/v1/config	shown by GET, patchable, with a model reload

The KV-RAM host tier’s budget: page-locked whole at start and held for the life of the load, so it is RAM the process holds even idle (and the figure Windows reports as its shared GPU memory). 0 disables the tier.

switch

switch.drain_timeout

Flag	--switch-drain-timeout
Env var	IGNIS_SWITCH_DRAIN_TIMEOUT
Kind	seconds
Default	30
Rule	0..=3600
Models	all
/v1/config	shown by GET, patchable live

How long a model switch lets the old model’s running requests finish before it cancels them, in seconds; 0 cancels them at once.

switch.allow_implicit

Flag	--switch-allow-implicit (may stand bare: true)
Env var	IGNIS_SWITCH_ALLOW_IMPLICIT
Kind	bool
Default	true
Models	all
/v1/config	shown by GET, patchable live

Whether a request whose model names another model known_models lists switches the server to it, then is served on it. Off refuses it as an unknown model is. An explicit switch is not affected.

switch.known_models

Flag	--switch-known-models
Env var	IGNIS_SWITCH_KNOWN_MODELS
Kind	id=path;..., repeatable
Default	{}
Models	all
/v1/config	shown by GET, patchable live

The models a request may switch to by naming them, each with the artifact it loads from: <id>=<path> (the flag is repeatable, the env var takes ;-separated pairs, a file a map). The model the server starts on joins them once its load says its id.

spec

spec.backend

Flag	--spec-backend
Env var	IGNIS_SPEC_BACKEND
Kind	backend
Default	unset
Rule	one of dflash2, mtp, off
Models	all
Per model	--qwen38-spec-backend, --qwen38flashnext-spec-backend
/v1/config	shown by GET, patchable, with a model reload

The speculative backend: dflash2 (the 27B’s drafter), mtp (Flash-Next’s head, with its companion container beside the artifact), or off. Unset, no speculation.

spec.draft_tokens

Flag	--spec-draft-tokens
Env var	IGNIS_SPEC_DRAFT_TOKENS
Kind	tokens
Default	unset
Rule	1..=7
Models	all
Per model	--qwen38-spec-draft-tokens, --qwen38flashnext-spec-draft-tokens
/v1/config	shown by GET, patchable, with a model reload

The draft window: required with dflash2; with mtp, the most drafts a lane verifies (default 2). Needs a backend.

spec.draft_rows

Flag	--spec-draft-rows
Env var	IGNIS_SPEC_DRAFT_ROWS
Kind	rows
Default	unset
Rule	0..=32
Models	Qwen3.8-Flash-Next
/v1/config	shown by GET, patchable, with a model reload

Flash-Next’s draft row budget: the rows a verify round takes across lanes, 0 (= 8) or 2..=32; past 8, five to eight lanes draft too.

spec.decode_lanes

Flag	--spec-decode-lanes
Env var	IGNIS_SPEC_DECODE_LANES
Kind	lanes
Default	unset
Rule	1..=8
Models	all
Per model	--qwen38-spec-decode-lanes, --qwen38flashnext-spec-decode-lanes
/v1/config	shown by GET, patchable, with a model reload

The decode lanes: the sequences a model decodes at once. On the 27B each lane reserves its state (~233 MB with dflash2 under hq-e8-2b) and its share of the round buffers, and takes two of the default host retained slots. On Flash-Next they decode sharing the KV pool — min(524,288 tokens, lanes x max_context), never below one max_context and a page per retained slot. Unset, the model’s own: 8 on the 27B, 3 on Flash-Next.

spec.draft_head

Flag	--spec-draft-head
Env var	IGNIS_SPEC_DRAFT_HEAD
Kind	head
Default	unset
Rule	one of full, shortlist, auto
Models	all
Per model	--qwen38-spec-draft-head, --qwen38flashnext-spec-draft-head
/v1/config	shown by GET, patchable, with a model reload

The drafter’s proposal head: full (the target’s output head), shortlist (the artifact’s Q4 head over the 131,072 most frequent tokens, ~357 MB more VRAM) or auto (full on every card class: the shortlist frees no VRAM and measured slower). Needs a backend; mtp has only full. Unset, full.

vision

vision.enabled

Flag	--vision-enabled (may stand bare: true)
Env var	IGNIS_VISION_ENABLED
Kind	bool
Default	false
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

Load the vision tower and reserve its workspace.

vision.max_tokens

Flag	--vision-max-tokens
Env var	IGNIS_VISION_MAX_TOKENS
Kind	tokens
Default	32768
Rule	1..=1048576
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

Merged vision tokens per request. An image over it is shrunk to fit, aspect kept; refused only when its aspect ratio cannot fit, or when several images total over it. Needs vision on.

vision.embedding_pool_mib

Flag	--vision-embedding-pool-mib
Env var	IGNIS_VISION_EMBEDDING_POOL_MIB
Kind	MiB
Default	unset
Rule	1..=65536
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

Encoded images kept for reuse, in MiB, where vision.residency puts the pool: VRAM, or pinned host RAM under host. Unset, one envelope-wide embedding; a pool below that is raised to it. Needs vision on.

vision.residency

Flag	--vision-residency
Env var	IGNIS_VISION_RESIDENCY
Kind	vision residency
Default	host
Rule	one of host, device
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

Where the vision tower’s weights and the embedding pool live: device, in VRAM for the life of the load (today’s encode latency), or host, in pinned host RAM, the weights copied to the GPU for each image encode (~300 MB and the pool less VRAM, so more KV pool, a few ms more an encode). A host start the host RAM cannot hold is refused, never moved to the device. Needs vision on.

media

media.allow_private_network

Flag	--media-allow-private-network (may stand bare: true)
Env var	IGNIS_MEDIA_ALLOW_PRIVATE_NETWORK
Kind	bool
Default	false
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

Fetch image URLs on private, loopback, link-local, multicast and CGNAT addresses. Off, so an exposed server cannot be used to probe the operator’s LAN. Needs vision on.

media.cache_mib

Flag	--media-cache-mib
Env var	IGNIS_MEDIA_CACHE_MIB
Kind	MiB
Default	1024
Rule	0..=65536
Models	Qwen3.8-27B
/v1/config	shown by GET, patchable, with a model reload

Prepared images kept for reuse, in MiB; 0 keeps none. Needs vision on.

image

image.queue_depth

Flag	--image-queue-depth
Env var	IGNIS_IMAGE_QUEUE_DEPTH
Kind	jobs
Default	4
Rule	0..=64
Models	all
/v1/config	shown by GET, patchable live

The image jobs that may wait behind the one running, in order; one more is refused 503 engine_full with Retry-After. 0 runs one at a time with none waiting.

image.max_n

Flag	--image-max-n
Env var	IGNIS_IMAGE_MAX_N
Kind	images
Default	10
Rule	1..=64
Models	all
/v1/config	shown by GET, patchable live

The most images one request may ask for (n), made one after another over the job’s one condition prefix.

image.max_condition_images

Flag	--image-max-condition-images
Env var	IGNIS_IMAGE_MAX_CONDITION_IMAGES
Kind	images
Default	10
Rule	1..=10
Models	all
/v1/config	shown by GET, patchable live

The most condition images one edit may send. At most 10, what the image model’s encoder holds.

image.max_condition_mipx

Flag	--image-max-condition-mipx
Env var	IGNIS_IMAGE_MAX_CONDITION_MIPX
Kind	Mipx
Default	16
Rule	1..=64
Models	all
/v1/config	shown by GET, patchable live

The largest a condition image may decode to, in Mi pixels (1,048,576 pixels, a 1024x1024 image, each); a larger one is refused. A condition image is encoded at 1024x1024 of area at most, so 16 — a 16-megapixel photo, 64 MiB decoded — takes any phone camera’s own output and bounds what one condition image decodes to.

image.max_condition_decodes

Flag	--image-max-condition-decodes
Env var	IGNIS_IMAGE_MAX_CONDITION_DECODES
Kind	jobs
Default	1
Rule	1..=16
Models	all
/v1/config	shown by GET, patchable live

The most edits that decode their condition images at once; the others wait their turn. Each holds up to max_condition_images frames of up to max_condition_mipx while it decodes, and the queue runs one job at a time, so more side by side buy memory, not throughput.

image.sizes

Flag	--image-sizes
Env var	IGNIS_IMAGE_SIZES
Kind	WxH;..., repeatable
Default	["2048x2048","2400x1792","1792x2400","2528x1696","1696x2528","2752x1536","1536x2752","1024x1024"]
Models	all
/v1/config	shown by GET, patchable live

The sizes a request may ask for, <width>x<height> with each side a multiple of 32: by default the checkpoint’s presets and 1024x1024, the size OpenAI clients often send. auto picks among them.

image.max_target_cells

Flag	--image-max-target-cells
Env var	IGNIS_IMAGE_MAX_TARGET_CELLS
Kind	cells
Default	16800
Rule	1..=1048576
Models	all
/v1/config	shown by GET, patchable live

The largest image the image model’s load makes room for, in latent cells of 16x16 pixels: 16,800, 2400x1792, the checkpoint’s largest preset. The load sizes its working memory for the largest of sizes within it; a size past it is not served, and the load warns. Read by the load, so a change holds from the next one.

image.condition_budget_tokens

Flag	--image-condition-budget-tokens
Env var	IGNIS_IMAGE_CONDITION_BUDGET_TOKENS
Kind	tokens
Default	0
Models	all
/v1/config	shown by GET, patchable live

The condition budget, in condition-prefix tokens (~256 KiB each, a 1024x1024 condition image 4,096 of them). 0 is the whole budget the load planned; a value can only lower it.

image.min_condition_budget_tokens

Flag	--image-min-condition-budget-tokens
Env var	IGNIS_IMAGE_MIN_CONDITION_BUDGET_TOKENS
Kind	tokens
Default	8191
Rule	0..=1048576
Models	all
/v1/config	shown by GET, patchable live

The least condition budget the image model’s load accepts, in condition-prefix tokens: a load whose VRAM plan leaves less is refused, naming the shortfall, rather than refuse its requests one by one. 8,191, the longest text-to-image prompt the image model’s positions admit; 0 takes any. Read by the load, so a change holds from the next one.

image.condition_floor_px

Flag	--image-condition-floor-px
Env var	IGNIS_IMAGE_CONDITION_FLOOR_PX
Kind	pixels
Default	512
Rule	32..=1024
Models	all
/v1/config	shown by GET, patchable live

The smallest the budget may scale a condition image down to, as the side of a square of the same area, in pixels; a request whose images would go below it is refused, naming the budget.

image.prompt_enhancement

Flag	--image-prompt-enhancement (may stand bare: true)
Env var	IGNIS_IMAGE_PROMPT_ENHANCEMENT
Kind	bool
Default	false
Models	all
/v1/config	shown by GET, patchable live

Whether a request that does not say has its prompt rewritten by prompt_enhancer before the image model reads it.

image.prompt_enhancer

Flag	--image-prompt-enhancer
Env var	IGNIS_IMAGE_PROMPT_ENHANCER
Kind	text
Default	qwen3.8-27b
Models	all
/v1/config	shown by GET, patchable live

The text model that rewrites prompts: the served id of a model an explicit switch may name (switch.known_models, the start model, or a downloaded catalog entry), whatever switch.allow_implicit says, or the text model loaded. A PATCH naming any other is refused, and a start warns of it; a request that meets one anyway skips the rewrite, with a warning.

image.prompt_enhancer_max_tokens

Flag	--image-prompt-enhancer-max-tokens
Env var	IGNIS_IMAGE_PROMPT_ENHANCER_MAX_TOKENS
Kind	tokens
Default	1024
Rule	16..=8192
Models	all
/v1/config	shown by GET, patchable live

The most tokens one rewrite may take. Qwen’s prompts ask for about 200 words; a longer answer is cut here and used as it is.

speech

speech.model

Flag	--speech-model
Env var	IGNIS_SPEECH_MODEL
Kind	text
Default	auto
Models	all
/v1/config	shown by GET, patchable live

The speech model: auto (the 1.7B Q4_K_M catalog entry, qwen3-asr-1.7b), a catalog entry’s id, or off. It is enabled when its files are on disk; nothing is ever downloaded by itself (ignis-server model download <id>).

speech.compat_names

Flag	--speech-compat-names
Env var	IGNIS_SPEECH_COMPAT_NAMES
Kind	name;..., repeatable
Default	["whisper-1","gpt-4o-transcribe","gpt-4o-mini-transcribe","gpt-transcribe","gpt-live-transcribe"]
Models	all
/v1/config	shown by GET, patchable live

The model names OpenAI clients send that the speech model answers to, so an unmodified client works: Aider hard-codes whisper-1. Not catalog aliases and not listed by GET /v1/models; a name equal to a served id is refused at start.

speech.threads

Flag	--speech-threads
Env var	IGNIS_SPEECH_THREADS
Kind	count|auto
Default	auto
Models	all
/v1/config	shown by GET, patchable live

The threads the speech model computes with: auto is the physical cores less one, counted within the CPUs the server may run on (a container’s cpuset or quota). Read by the load, so a change holds from the next one; each load logs the count (ignis.speech.threads).

speech.max_clip_seconds

Flag	--speech-max-clip-seconds
Env var	IGNIS_SPEECH_MAX_CLIP_SECONDS
Kind	seconds
Default	1200
Rule	1..=1200
Models	all
/v1/config	shown by GET, patchable live

The longest clip one transcription takes, in seconds; a longer one is refused and nothing is split. 1,200, the model’s limit.

speech.context_max_tokens

Flag	--speech-context-max-tokens
Env var	IGNIS_SPEECH_CONTEXT_MAX_TOKENS
Kind	tokens
Default	512
Rule	0..=4096
Models	all
/v1/config	shown by GET, patchable live

The most tokens of context (prompt and keywords[]) a transcription reads; over it the last ones are kept. 0 reads none.

speech.context_copy_words

Flag	--speech-context-copy-words
Env var	IGNIS_SPEECH_CONTEXT_COPY_WORDS
Kind	words
Default	3
Rule	0..=32
Models	all
/v1/config	shown by GET, patchable live

On a pause the model can write a line of its context in place of what it heard: a transcript that only repeats this many words in a row of one context line, or a whole line of two words or more, is taken for no speech and comes back empty. One word alone never is. 0 turns the guard off. Read when a session opens, so a change holds from the next one.

speech.pinned_weights

Flag	--speech-pinned-weights (may stand bare: true)
Env var	IGNIS_SPEECH_PINNED_WEIGHTS
Kind	bool
Default	false
Models	all
/v1/config	shown by GET, patchable live

Pin the speech model’s weights in host memory instead of mapping the file. Read by the load, so a change holds from the next one.

speech.idle_unload

Flag	--speech-idle-unload
Env var	IGNIS_SPEECH_IDLE_UNLOAD
Kind	seconds
Default	300
Rule	0..=604800
Models	all
/v1/config	shown by GET, patchable live

Seconds with no transcription, and no Realtime session open, after which the speech model is unloaded, its host memory given back; the next transcription loads it again. 0 never unloads.

speech.partial_interval_ms

Flag	--speech-partial-interval-ms
Env var	IGNIS_SPEECH_PARTIAL_INTERVAL_MS
Kind	ms
Default	2000
Rule	100..=60000
Models	all
/v1/config	shown by GET, patchable live

How often at most a Realtime session’s live transcript is refreshed while audio arrives, in milliseconds: each refresh sends the new stable text, and the provisional tail to a session that opted in. Best effort: a refresh is skipped while the session’s windows are still being computed, and waits longer after a costly one (speech.partial_cost_factor). Read by every open session at its next refresh.

speech.partial_cost_factor

Flag	--speech-partial-cost-factor
Env var	IGNIS_SPEECH_PARTIAL_COST_FACTOR
Kind	times
Default	1
Rule	1..=16
Models	all
/v1/config	shown by GET, patchable live

How long a Realtime session’s next refresh waits at least, as a multiple of the compute time of its last one, from when that one was queued; never less than speech.partial_interval_ms. Each refresh decodes the item so far again, so it costs more as the item grows: the refreshes then take at most about one part in this many of the speech worker. 1 waits only for the last refresh to finish; 2 and 3 measured no faster commits on a 10-core or a 4-core CPU. Read by every open session at its next refresh.

speech.rollback_tokens

Flag	--speech-rollback-tokens
Env var	IGNIS_SPEECH_ROLLBACK_TOKENS
Kind	tokens
Default	10
Rule	0..=64
Models	all
/v1/config	shown by GET, patchable live

The tokens of a live transcript held back as provisional, the rest being stable: the tail a Realtime session that asks for it gets. Read by the speech executor when a session opens, so a change holds from the next session.

speech.unfixed_partials

Flag	--speech-unfixed-partials
Env var	IGNIS_SPEECH_UNFIXED_PARTIALS
Kind	partials
Default	2
Rule	0..=16
Models	all
/v1/config	shown by GET, patchable live

The first refreshes of a Realtime session’s live transcript that make no text stable, whatever speech.rollback_tokens says: the language and the opening words are decided on more than one interval of audio (the official streaming’s unfixed_chunk_num). 0 fixes text from the first refresh. Read by the speech executor when a session opens, so a change holds from the next session.

speech.max_sessions

Flag	--speech-max-sessions
Env var	IGNIS_SPEECH_MAX_SESSIONS
Kind	sessions
Default	4
Rule	1..=64
Models	all
/v1/config	shown by GET, patchable live

The most Realtime transcription sessions open at once; past it the next is refused at the upgrade 503 with Retry-After.

speech.max_queue

Flag	--speech-max-queue
Env var	IGNIS_SPEECH_MAX_QUEUE
Kind	requests
Default	4
Rule	0..=64
Models	all
/v1/config	shown by GET, patchable live

The batch transcriptions that may wait behind the one running, in order; one more is refused 503 engine_full with Retry-After. 0 runs one at a time with none waiting.

speech.vad_threshold

Flag	--speech-vad-threshold
Env var	IGNIS_SPEECH_VAD_THRESHOLD
Kind	0..1
Default	0.5
Models	all
/v1/config	shown by GET, patchable live

The speech probability at or over which server_vad hears speech in 32 ms of a Realtime session’s audio: the default of its threshold. Read when a session opens.

speech.vad_prefix_padding_ms

Flag	--speech-vad-prefix-padding-ms
Env var	IGNIS_SPEECH_VAD_PREFIX_PADDING_MS
Kind	ms
Default	300
Rule	0..=10000
Models	all
/v1/config	shown by GET, patchable live

The audio from before the speech that a server_vad item keeps, in milliseconds: the default of its prefix_padding_ms. Read when a session opens.

speech.vad_silence_duration_ms

Flag	--speech-vad-silence-duration-ms
Env var	IGNIS_SPEECH_VAD_SILENCE_DURATION_MS
Kind	ms
Default	500
Rule	0..=60000
Models	all
/v1/config	shown by GET, patchable live

The silence after the speech at which server_vad commits the item, in milliseconds: the default of its silence_duration_ms. Read when a session opens.

speech.vad_leading_silence_ms

Flag	--speech-vad-leading-silence-ms
Env var	IGNIS_SPEECH_VAD_LEADING_SILENCE_MS
Kind	ms
Default	500
Rule	0..=10000
Models	all
/v1/config	shown by GET, patchable live

The digital silence put before each server_vad item’s audio when it is fed to the speech model, in milliseconds: the model’s transcript depends on the audio before the speech, and an item starts only its prefix padding before it. No event and no usage counts it; it counts toward speech.max_clip_seconds, as the model reads it. A manual commit’s item gets none. Read at each item’s start.

speech.loan_grace_seconds

Flag	--speech-loan-grace-seconds
Env var	IGNIS_SPEECH_LOAN_GRACE_SECONDS
Kind	seconds
Default	30
Rule	0..=3600
Models	all
/v1/config	shown by GET, patchable live

How long a loan of free KV pages to the speech model outlives its transcription, in seconds (spec speech/06), not served yet.

speech.loan_wait_seconds

Flag	--speech-loan-wait-seconds
Env var	IGNIS_SPEECH_LOAN_WAIT_SECONDS
Kind	seconds
Default	5
Rule	0..=600
Models	all
/v1/config	shown by GET, patchable live

How long a transcription waits for a loan before it runs on the CPU, in seconds (spec speech/06), not served yet.

speech.max_utterance_seconds

Flag	--speech-max-utterance-seconds
Env var	IGNIS_SPEECH_MAX_UTTERANCE_SECONDS
Kind	seconds
Default	120
Rule	0..=1200
Models	all
/v1/config	shown by GET, patchable live

The length guard: once a server_vad item holds this many seconds of audio, the server ends it at the next 32 ms the VAD hears under its threshold, a pause too short to end it, or speech.utterance_cut_search_ms later at the latest, and the speech goes on in a new item; so a room that never falls silent loses nothing to speech.max_clip_seconds. A session’s own session.ignis.max_utterance_seconds overrides it. 0 never cuts. Read at each item’s audio.

speech.utterance_cut_search_ms

Flag	--speech-utterance-cut-search-ms
Env var	IGNIS_SPEECH_UTTERANCE_CUT_SEARCH_MS
Kind	ms
Default	10000
Rule	0..=60000
Models	all
/v1/config	shown by GET, patchable live

How long past speech.max_utterance_seconds the length guard looks for a pause, in milliseconds, before it cuts where the audio is. The search looks forward only, so no text already sent is past the cut. In fluent speech the VAD hears a pause within 10 s 95 times in 100 (spec speech/08, As built).

speech.max_queued_audio_seconds

Flag	--speech-max-queued-audio-seconds
Env var	IGNIS_SPEECH_MAX_QUEUED_AUDIO_SECONDS
Kind	seconds
Default	120
Rule	1..=3600
Models	all
/v1/config	shown by GET, patchable live

The audio a Realtime session may hold that is not computed yet, in seconds: received, and still waiting for the speech model’s windows or the diarizer’s steps. An append past it is refused with an error, ignis_audio_queue_full, and the session stays open: a client feeding a file faster than the CPU computes it paces itself by it, and one fed at real time never reaches it. Read at each append.

diarizer

diarizer.model

Flag	--diarizer-model
Env var	IGNIS_DIARIZER_MODEL
Kind	text
Default	auto
Models	all
/v1/config	shown by GET, patchable live

The diarizer: auto (the catalog’s diarizer entry), a catalog entry’s id, or off. It is enabled when its files are on disk; nothing is ever downloaded by itself. A text, image or speech model’s id is refused. Loaded by the first session that asks for speakers, unloaded with the speech model.

diarizer.threads

Flag	--diarizer-threads
Env var	IGNIS_DIARIZER_THREADS
Kind	count|auto
Default	auto
Models	all
/v1/config	shown by GET, patchable live

The threads the diarizer computes with, a pool of its own beside the speech model’s, so its steps never wait for the speech model’s work: auto is a third of the physical cores, at least 2 (3 on a 10-core machine), counted within the CPUs the server may run on. Read by the load, so a change holds from the next one; the load logs the count (ignis.diarizer.load).

diarizer.min_speaker_seconds

Flag	--diarizer-min-speaker-seconds
Env var	IGNIS_DIARIZER_MIN_SPEAKER_SECONDS
Kind	number
Default	2.0
Rule	0..=60
Models	all
/v1/config	shown by GET, patchable live

The ghost rule, in seconds of speech: a voice the diarizer hears is a speaker of its own once it has spoken this long, and until then its speech counts for the speaker it sounds most like, so a short mistaken voice splits no utterance. The session’s first voice is a speaker at once. 0 makes every voice a speaker at its first frame. At 2 s a ghost voice of the synthetic conversations goes, and no real meeting’s quiet participant is lost (spec speech/08, As built).

diarizer.live_speaker_cache_frames

Flag	--diarizer-live-speaker-cache-frames
Env var	IGNIS_DIARIZER_LIVE_SPEAKER_CACHE_FRAMES
Kind	frames
Default	264
Rule	8..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The live geometry’s speaker cache, in 80 ms frames: what the diarizer keeps of every speaker heard so far.

diarizer.live_fifo_frames

Flag	--diarizer-live-fifo-frames
Env var	IGNIS_DIARIZER_LIVE_FIFO_FRAMES
Kind	frames
Default	264
Rule	0..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The live geometry’s FIFO, in 80 ms frames: the latest frames each step reads beside its chunk.

diarizer.live_chunk_frames

Flag	--diarizer-live-chunk-frames
Env var	IGNIS_DIARIZER_LIVE_CHUNK_FRAMES
Kind	frames
Default	25
Rule	1..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The live geometry’s chunk, in 80 ms frames: the frames one step scores. 25 is 2 s.

diarizer.live_right_context_frames

Flag	--diarizer-live-right-context-frames
Env var	IGNIS_DIARIZER_LIVE_RIGHT_CONTEXT_FRAMES
Kind	frames
Default	4
Rule	0..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The live geometry’s right context, in 80 ms frames: the look-ahead a step waits for. With the chunk, the latency: 2.32 s at 25 and 4.

diarizer.live_update_period_frames

Flag	--diarizer-live-update-period-frames
Env var	IGNIS_DIARIZER_LIVE_UPDATE_PERIOD_FRAMES
Kind	frames
Default	222
Rule	1..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The live geometry’s update period, in 80 ms frames: the oldest frames an overflowing FIFO gives the speaker cache at once.

diarizer.file_speaker_cache_frames

Flag	--diarizer-file-speaker-cache-frames
Env var	IGNIS_DIARIZER_FILE_SPEAKER_CACHE_FRAMES
Kind	frames
Default	264
Rule	8..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The file geometry’s speaker cache, in 80 ms frames, for a session fed faster than real time (session.ignis.diarization: "file").

diarizer.file_fifo_frames

Flag	--diarizer-file-fifo-frames
Env var	IGNIS_DIARIZER_FILE_FIFO_FRAMES
Kind	frames
Default	40
Rule	0..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The file geometry’s FIFO, in 80 ms frames.

diarizer.file_chunk_frames

Flag	--diarizer-file-chunk-frames
Env var	IGNIS_DIARIZER_FILE_CHUNK_FRAMES
Kind	frames
Default	340
Rule	1..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The file geometry’s chunk, in 80 ms frames: 340 is 27.2 s.

diarizer.file_right_context_frames

Flag	--diarizer-file-right-context-frames
Env var	IGNIS_DIARIZER_FILE_RIGHT_CONTEXT_FRAMES
Kind	frames
Default	40
Rule	0..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The file geometry’s right context, in 80 ms frames: with the chunk, 30.4 s of latency at 340 and 40.

diarizer.file_update_period_frames

Flag	--diarizer-file-update-period-frames
Env var	IGNIS_DIARIZER_FILE_UPDATE_PERIOD_FRAMES
Kind	frames
Default	300
Rule	1..=1000
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The file geometry’s update period, in 80 ms frames.

diarizer.prediction_score_threshold

Flag	--diarizer-prediction-score-threshold
Env var	IGNIS_DIARIZER_PREDICTION_SCORE_THRESHOLD
Kind	number|auto
Default	auto
Rule	0.001..=1
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The speaker cache’s floor on a frame’s probabilities when it scores the frames to keep. auto is the checkpoint’s (0.25).

diarizer.latest_frames_score_boost

Flag	--diarizer-latest-frames-score-boost
Env var	IGNIS_DIARIZER_LATEST_FRAMES_SCORE_BOOST
Kind	number|auto
Default	auto
Rule	0..=10
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

What the speaker cache adds to the scores of the frames the FIFO just gave it. auto is the checkpoint’s (0.05).

diarizer.strong_boost_rate

Flag	--diarizer-strong-boost-rate
Env var	IGNIS_DIARIZER_STRONG_BOOST_RATE
Kind	number|auto
Default	auto
Rule	0..=4
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The share of each speaker’s room in the speaker cache whose best frames get the strong boost. auto is the checkpoint’s (0.75).

diarizer.weak_boost_rate

Flag	--diarizer-weak-boost-rate
Env var	IGNIS_DIARIZER_WEAK_BOOST_RATE
Kind	number|auto
Default	auto
Rule	0..=4
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The share of each speaker’s room in the speaker cache whose best frames get the weak boost. auto is the checkpoint’s (1.5).

diarizer.min_positive_scores_rate

Flag	--diarizer-min-positive-scores-rate
Env var	IGNIS_DIARIZER_MIN_POSITIVE_SCORES_RATE
Kind	number|auto
Default	auto
Rule	0..=1
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The share of each speaker’s room a speaker’s positive scores must fill before its other frames stop counting. auto is the checkpoint’s (0.5).

diarizer.speaker_cache_silence_frames_per_speaker

Flag	--diarizer-speaker-cache-silence-frames-per-speaker
Env var	IGNIS_DIARIZER_SPEAKER_CACHE_SILENCE_FRAMES_PER_SPEAKER
Kind	count|auto
Default	auto
Rule	0..=125
Models	Nemotron 3 Diarization
/v1/config	shown by GET, patchable live

The silence frames the speaker cache keeps for every speaker. auto is the checkpoint’s (1).

download

download.enabled

Flag	--download-enabled (may stand bare: true)
Env var	IGNIS_DOWNLOAD_ENABLED
Kind	bool
Default	true
Models	all
/v1/config	shown by GET, not patchable

Fetch the model when no artifact is on disk — asked first when stdin is a terminal, downloaded straight away when it is not. Off keeps the placeholder template, and refuses POST /v1/models/download too: the server never uses the network.

download.path

Flag	--download-path
Env var	IGNIS_DOWNLOAD_PATH
Kind	path
Default	./models
Models	all
/v1/config	shown by GET, not patchable

Where a fetched model lands, and where one fetched earlier is found.

download.endpoint

Flag	--download-endpoint
Env var	IGNIS_DOWNLOAD_ENDPOINT
Kind	text
Default	https://huggingface.co
Models	all
/v1/config	shown by GET, not patchable

Where a fetched model’s files come from: Hugging Face, or a mirror answering its {repo}/resolve/{revision}/{file} route (an Artifactory or Nexus proxy, a static server laid out the same way). The pins still say whether the bytes are the published ones.

download.token

Flag	--download-token
Env var	IGNIS_DOWNLOAD_TOKEN
Kind	token
Default	unset
Models	all
/v1/config	hidden from GET, not patchable

Sent as Authorization: Bearer to the endpoint, never across a redirect to another host. Unset, HF_TOKEN is sent to huggingface.co and nothing to a mirror. Never shown by GET /v1/config or config print.

download.catalog

Flag	--download-catalog
Env var	IGNIS_DOWNLOAD_CATALOG
Kind	path
Default	unset
Models	all
/v1/config	shown by GET, not patchable

An operator catalog: a YAML or JSON file of models fetched beside the built-in ones, in the same format; an id the built-in catalog has is refused. A relative path is read from the directory of the config file that names it.

download.from_ui

Flag	--download-from-ui (may stand bare: true)
Env var	IGNIS_DOWNLOAD_FROM_UI
Kind	bool
Default	true
Models	all
/v1/config	shown by GET, not patchable

Let POST /v1/models/download, the Server view’s Download, fetch a catalog entry while the server runs, under the same API key as PATCH /v1/config; it needs enabled too. Off refuses it, and model download on the server’s machine still fetches. Off refuses POST /v1/models/remove too, the Server view’s Remove.

download.free_margin

Flag	--download-free-margin
Env var	IGNIS_DOWNLOAD_FREE_MARGIN
Kind	bytes
Default	1G
Models	all
/v1/config	shown by GET, not patchable

What a download must leave free on the volume of path beyond the bytes it still misses: the Server view’s Download, model download and the start’s own fetch are refused, before any byte is written, when the free space is below the two together.

ngram

ngram.persist

Flag	--ngram-persist (may stand bare: true)
Env var	IGNIS_NGRAM_PERSIST
Kind	bool
Default	true
Models	all
/v1/config	shown by GET, patchable, with a model reload

Persist Flash-Next’s n-gram hot rows between loads.

ngram.persist_path

Flag	--ngram-persist-path
Env var	IGNIS_NGRAM_PERSIST_PATH
Kind	model|auto|dir
Default	model
Models	all
/v1/config	shown by GET, patchable, with a model reload

Where the n-gram cache goes: model (beside the artifact), auto (Windows LOCALAPPDATA/ignis/cache/ngram, Linux XDG_CACHE_HOME/ignis/ngram or HOME/.cache/ignis/ngram), or a directory.

ngram.hot_bytes

Flag	--ngram-hot-bytes
Env var	IGNIS_NGRAM_HOT_BYTES
Kind	bytes|auto
Default	unset
Models	Qwen3.8-Flash-Next
/v1/config	shown by GET, patchable, with a model reload

The n-gram rows held in RAM, every other row read from NVMe when a step needs it: a byte count, or auto — what the host plan leaves after its other lines and the margin, less 256 MiB for the load, the whole ~29 GB table when that fits, never below the default. Unset, 1 GiB.

kv_disk

kv_disk.bytes

Flag	--kv-disk-bytes
Env var	IGNIS_KV_DISK_BYTES
Kind	bytes
Default	unset
Models	all
Per model	--qwen38-kv-disk-bytes, --qwen38flashnext-kv-disk-bytes
/v1/config	shown by GET, patchable, with a model reload

KV-disk’s budget: a ceiling, cut at start to the volume’s free space less 10 GiB. Unset, the model’s own: 4G on Flash-Next, 0 (off) on the 27B.

kv_disk.path

Flag	--kv-disk-path
Env var	IGNIS_KV_DISK_PATH
Kind	model|auto|dir
Default	model
Models	all
/v1/config	shown by GET, patchable, with a model reload

Where KV-disk’s files go: model (beside the artifact), auto (the per-user cache directory under kv-disk), or a directory. They go in an ignis-kv-disk/- directory there, removed at shutdown.