Metrics

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.

The metrics listener (on by default, keyless, on loopback) serves these in the Prometheus text format, and the Playground’s Monitor reads the same set at /ui/metrics (Telemetry and metrics). A family marked always is in every scrape from the start, zeros included; the others appear once what the last column names has happened, and then stay.

Metric	Type	Meaning	Exported
ignis_build_info	gauge	Constant build identity with value 1.	always
ignis_condition_budget_used_bytes	gauge	Bytes of the condition prefixes the image model’s jobs hold now.	always
ignis_decision_answer_mass	histogram	Share of the next-token distribution held by a decision’s declared options.	after the first decision
ignis_decisions_total	counter	Questions answered by a readout, by typed primitive.	after the first decision
ignis_decoded_tokens_total	counter	Tokens generated so far, counted as each one is emitted.	always
ignis_expert_bytes_moved_total	counter	Bytes of expert projections copied host-to-device, misses and prefetches, by phase.	on a Flash-Next load
ignis_expert_cache_hits_total	counter	Selected expert projections already resident in the VRAM expert cache, or staged for their layer, by K class and phase.	on a Flash-Next load
ignis_expert_cache_misses_total	counter	Selected expert projections copied in from the pinned host pool by their own step, by K class and phase.	on a Flash-Next load
ignis_expert_cache_slots	gauge	VRAM expert cache LRU slots per K class: capacity reserved at load, and in use. A pinned hot set’s slots are not among them (the load’s ignis.runtime.flash_next_pinned_experts event names them).	on a Flash-Next load
ignis_expert_prefetches_issued_total	counter	Expert projections copied ahead for the next layer by the router lookahead.	on a Flash-Next load
ignis_expert_prefetches_used_total	counter	Prefetched expert projections at their first use, whenever it comes.	on a Flash-Next load
ignis_expert_residency_stall_seconds_total	counter	Device time the expert kernels waited on their step’s demand copies (the misses), by phase; a step with no miss adds nothing.	on a Flash-Next load
ignis_generated_tokens_total	counter	Generated tokens on completed requests.	always
ignis_image_images_total	counter	Images the image jobs made.	on a server that has seen an image job
ignis_image_job_duration_seconds	histogram	Image job duration from its turn to its last image, completed jobs only.	on a server that has seen an image job
ignis_image_job_phase_seconds_total	counter	Seconds the image jobs spent in each phase: waiting for their turn, encoding the condition, denoising, decoding.	on a server that has seen an image job
ignis_image_jobs_total	counter	Image jobs ended, by outcome.	on a server that has seen an image job
ignis_image_queue_jobs	gauge	Image jobs waiting for their turn and running (at most one).	on a server that has seen an image job
ignis_kv_cache_evictions_total	counter	Cumulative host-tier evictions.	always
ignis_kv_disk_bytes	gauge	The KV-disk tier: its effective budget, and the bytes its files hold.	on a load with the KV-disk tier
ignis_kv_disk_failures_total	counter	KV-disk writes refused or failed (nothing lost), and files that failed their check on the way back.	on a load with the KV-disk tier
ignis_kv_disk_spills_total	counter	Live snapshots written to KV-disk, by the tier they came from.	on a load with the KV-disk tier
ignis_kv_page_bytes	gauge	One KV page’s bytes.	always
ignis_kv_pool_pages	gauge	Pages the KV pool holds.	always
ignis_kv_pool_used_pages	gauge	KV pool pages reserved by running requests and retained state.	always
ignis_kv_ram_arena_bytes	gauge	The pinned host KV-RAM arena: what it holds, and what is used of it.	always
ignis_kv_ram_evictions_total	counter	Live host-tier snapshots dropped from KV-RAM to make room; the request re-prefills from the start.	always
ignis_locates_total	counter	Locate questions answered, by resolved kind, method, compression and whether they found an answer.	after the first locate
ignis_loop_guard_looped_requests_total	counter	Watched requests with at least one loop.	once the loop guard has watched a request
ignis_loop_guard_loops_total	counter	Loops the guard saw, by what it did about them.	once the loop guard has watched a request
ignis_loop_guard_min_ratio	histogram	Lowest loop score each watched request reached.	once the loop guard has watched a request
ignis_loop_guard_rewound_tokens_total	counter	Generated tokens the guard’s rewinds took back.	once the loop guard has watched a request
ignis_loop_guard_watched_requests_total	counter	Requests the loop guard watched.	once the loop guard has watched a request
ignis_ngram_read_bytes_total	counter	Bytes of the n-gram reads issued to the artifact file.	on a Flash-Next load
ignis_ngram_reads_total	counter	Reads issued to the artifact file for n-gram rows the hot-row cache did not hold.	on a Flash-Next load
ignis_ngram_rows_total	counter	N-gram embedding rows staged for prefill spans and decode rounds, by source: the RAM hot-row cache or the artifact file.	on a Flash-Next load
ignis_prefix_reused_tokens_total	counter	Cumulative tokens skipped through sibling-prefix reuse.	always
ignis_request_duration_seconds	histogram	Submission-to-completion latency.	always
ignis_request_ttft_seconds	histogram	Submission-to-first-token latency.	always
ignis_requests_accepted_total	counter	Accepted submissions.	always
ignis_requests_cancelled_total	counter	Accepted requests cancelled before completion.	always
ignis_requests_completed_total	counter	Completed requests.	always
ignis_requests_rejected_total	counter	Rejected submissions by fixed reason.	always
ignis_responses_queued_requests	gauge	Responses WebSocket requests waiting in the server-wide admission queue.	always
ignis_responses_sockets	gauge	Open Responses API WebSocket connections.	always
ignis_retained_host_bytes	gauge	The pinned host block holding the host retained slots’ images.	always
ignis_retained_host_slots	gauge	Of the retained slots, those whose images live in the pinned host block.	always
ignis_retained_reused_tokens_total	counter	Cumulative tokens skipped through retained state, by residency tier.	always
ignis_retained_slot_skips_total	counter	Publishes and captures that found no retained slot or no tail page.	always
ignis_retained_slots	gauge	Retained slots this load hands out, and how many hold an image.	always
ignis_retained_state_discards_total	counter	Retained checkpoints and prefixes discarded from the tier.	always
ignis_retained_state_hits_total	counter	Retained state chosen to resume from or brought back, by residency tier.	always
ignis_retained_state_misses_total	counter	First prefill chunks with no retained checkpoint matching in the tier.	always
ignis_retained_state_restores_total	counter	Retained state restored from the tier.	always
ignis_retained_state_spills_total	counter	Retained checkpoints and prefixes spilled into the tier.	always
ignis_scheduler_requests	gauge	Current requests by observable scheduler state.	always
ignis_speculative_accepted_tokens_total	counter	Draft tokens committed.	after the first request that ran speculative verify rounds
ignis_speculative_drafted_tokens_total	counter	Draft tokens verified.	after the first request that ran speculative verify rounds
ignis_speculative_position_accepted_total	counter	Verify rounds that committed the draft at the position.	after the first request that ran speculative verify rounds
ignis_speculative_position_drafted_total	counter	Verify rounds that proposed a draft at the position.	after the first request that ran speculative verify rounds
ignis_speculative_rounds_total	counter	Verify rounds run.	after the first request that ran speculative verify rounds
ignis_speech_audio_seconds_total	counter	Seconds of audio the completed transcriptions covered.	on a server that has been asked for a transcription
ignis_speech_commit_wait_seconds	histogram	Wait from a transcription’s commit (a batch request’s last byte received) to its transcript, completed ones only.	on a server that has been asked for a transcription
ignis_speech_queue_length	gauge	Batch transcriptions waiting behind the one running.	on a server that has been asked for a transcription
ignis_speech_sessions_open	gauge	Realtime transcription sessions open.	on a server that has been asked for a transcription
ignis_speech_transcriptions_total	counter	Transcriptions ended, by executor and outcome.	on a server that has been asked for a transcription
ignis_thinking_forced_closes_total	counter	Completed requests whose reasoning block the thinking budget closed.	once the thinking budget has closed a reasoning block
ignis_tokenize_requests_total	counter	Calls to the counting routes, which never enter the scheduler, by route.	after the first call to a counting route
ignis_vram_budget_bytes	gauge	The device budget the plan was laid out inside.	always
ignis_vram_reserved_bytes	gauge	Device bytes this load reserved, by the plan line that reserved them.	always