Panel Instances

Panel Instances

The Instances and runtime workflows screen offers per-instance device selection, creation, launch parameters, reusable profiles, memory placement and guarded restarts. Open /rust/instances/workflows to see the current inventory and the action forms. Saved parameters do not necessarily equal the running launch configuration; check the running state and restart only when appropriate.

Main: select the serving engine

main is the pre-existing host service, not a deletable named instance. By default it uses the existing llama.cpp system unit and its saved parameters. A second, private settings file supports LexiEngine at panel/main-lexi-engine/params.env, with the selected engine persisted at panel/main-engine. Switching back preserves the llama.cpp scripts and parameters; the selected runtime determines main's launch plan, process status, log source, health and stop/restart controls.

The authenticated admin-only POST /api/instance/update route accepts {"id":"main","engine":"lexi-engine","params":{...}}, where params contains LexiEngine setting names such as MODEL, MMPROJ, BACKEND, DEVICES, CTX, CACHE_TYPE_K, CACHE_TYPE_V, KV_VULKAN_DTYPE, IMAGE_MIN_TOKENS, SPEC_TYPE, SPEC_DRAFT_MAX, HOST, PORT and API_KEY. For example:

{
  "id": "main",
  "engine": "lexi-engine",
  "devices": ["0000:01:00.0"],
  "params": {
    "MODEL": "/models/model.gguf",
    "MMPROJ": "/models/projector.gguf",
    "BACKEND": "vulkan",
    "HOST": "127.0.0.1",
    "PORT": 8081,
    "CTX": 200000,
    "PARALLEL": 1,
    "KV_VULKAN_DTYPE": "Q4_1",
    "IMAGE_MIN_TOKENS": 1024,
    "SPEC_TYPE": "mtp",
    "SPEC_DRAFT_MAX": 2
  }
}

Replace the example paths and PCI address with the installed files and reported device. A bare {"id":"main","engine":"llama.cpp"} switches back to the old service and its original parameters. To change the GPU selection, first stop main and submit the top-level devices array, or use the existing Change devices control. The params.DEVICES value, if supplied, must agree with the pinned selection. Legacy parameter edits remain on the Parameters page. All POST requests remain subject to the Panel's existing authenticated-session, same-origin and CSRF controls; do not send a bare unauthenticated HTTP request.

Safety and observability: The switch stops the previous managed service before selecting or starting the replacement. An unsuccessful start restores the previous engine selection and attempts to restart its service. Main's new launch logs use the same native launcher log path as other instances; /api/instances reports engine, unit, scope, pid, running, service_status and, when applicable, foreign_pid. A process the Panel did not start can be shown as served by another process (PID …), but Stop never adopts that PID. Native memory admission compares measured free VRAM or RAM with the GGUF shard size, projector and header-estimated KV allocation and rejects an insufficient card with required/free MiB figures. Memory required at the first actual inference request may exceed that lower bound; the native Engine's own admission checks still apply. When free memory cannot be measured, the switch is refused rather than trusting total device capacity.

The installed LexiEngine must report the requested projector and image-token flags. For remote clients, bind the engine to loopback and provide an appropriate authenticated TLS proxy; exposing native cleartext HTTP directly on a LAN is refused by the Engine. The legacy launcher is not rewritten to emulate these flags.

Current classic UI limitation: The API returns engine_choices for main, but the existing static classic Instances script does not yet render an engine-selector field; the separate JavaScript page is outside this update's allowed file ownership. The endpoint works independently of that missing visual picker. This implementation has not been qualified against a physical accelerator.

Serve a model on the selected GPU or CPU

1. Open Instances and runtime workflows, then Create and preview. Pick a reported GPU or CPU only, provide a new lowercase instance ID, optionally select the installed engine and backend, then provide a verified local model path and click Preview auto-fit. A preview does not start service and can refuse when a required signed runtime receipt or compatible build is unavailable.
2. Click Create instance only after inspecting device memory requirements and acknowledging the confirmation. The creation handler validates the installed engine and filesystem state.
3. Open Instance controls, select the newly created instance, then use Launch parameters to select an installed model, context and cache budget. Review the Saved versus running settings section; no restart occurs merely because you saved settings.
4. Start the selected instance from the instance controls. Verify actual health and memory admission before sending production requests.
5. To change device placement, stop the instance and use Select CPU and GPUs, choosing any reported GPU subset or CPU only, then Save device selection. Multiple GPUs require compatible backend support; CPU is available for host work without selecting it alongside a GPU.

The Serve a model on your chosen GPU or CPU guided task at /rust/instances/guides/instances.serve-on-device links these screens, checks actual inventory and create responses, and offers explicit reversal of creation/start steps. A new instance is not automatically selected in a later request.

Reusable instance profiles and launch presets

1. Choose the intended instance in Instance controls.
2. Under Reusable profiles, enter a lowercase name in Profile name, then click Save current instance profile. The saved snapshot is separate from a runtime launch.
3. After checking which model, devices, parameters and fallback tiers you intend to replace, enter that name and use Apply profile settings and fallback tiers.
4. Open Launch parameters or the native engine settings to inspect the saved values. Applying a profile need not restart a running engine; plan a guarded restart if active configuration must change.

The Launch parameters page at /rust/parameters explains each saved parameter beside its field and exposes built-in presets and fallback tiers. Invalid or unsupported flags are rejected by their existing parameter owners.

Parameters: edit, import, and review without automatic restart

Open Parameters at /rust/parameters for the saved values, parameter descriptions, engine capability checks, model profiles and fallback tiers. Saved versus running settings explains when live settings differ; use Instance controls and the guarded restart workflow to activate a saved change. The owner validates the save; no form silently starts inference.

To translate an existing launch command, use Translate a launch command → Launch command line → Preview command parameters. The preview calls POST /api/params/from-cmdline and displays proposed values and unrecognized flags. It does not save the suggestions, select a model or start service. Review them manually in Parameters, then explicitly save if appropriate. Operator and administrator roles may request this non-mutating preview; only administrators can save parameter changes. Command text is limited to 20,000 characters and malformed quoting is refused.

API key is intentionally a blank password field when editing. Leaving it blank preserves the existing key; choosing Remove saved API key clears it. Do not choose both a replacement and removal. The read-only table reports only whether a key is configured; it never displays its value. Model-profile results and parameter save notices also suppress the key. Saving a named profile omits the API key entirely, so credentials stay managed with the selected instance. Readers may inspect ordinary saved values but cannot submit changes; administrators alone see the write controls.

Templates: select, extract, save and download

Open Templates at /rust/templates. Installed templates lists each file, parser status, digest pin, preview and current selection. To select a valid template, choose Use next to it. The selected template passes through the parameter owner (POST /api/params); no inference restart is implied.

To create a template, enter Save as and Template body, then choose Validate and save template. For a template embedded in a model, use Use a model's own template to extract it through the model-aware owner. To fetch a direct .jinja link, use Download a template with URL and expected checksum or byte length when required. This follows the verified download owner, not an unverified shell command. Refresh the inventory after extraction or download. These actions are available only to administrators; readers see the inventory with no selection, save, extraction or download forms.

Restart safely, with explicit caveats

1. Open Safe restart and drain from /rust/instances/workflows for the selected instance.
2. Set Drain seconds (0–3600) and Gateway hold seconds (0–900). Enable Verify replacement before release; enable Attempt compatible cache handoff only when the running engine can support it.
3. Select the confirmation checkbox and click Save restart safeguards.
4. Select the confirmation checkbox and click Start guarded safe restart.
5. Open Inspect restart history and outcomes. A start acknowledgment is not a successful completion. Read the restart record and verify actual health.

A hold reduces the chance of new requests arriving mid-restart; a bounded drain gives in-flight requests time to finish. Neither step guarantees uninterrupted connections: long requests, client timeouts, failed readiness checks and unavailable replacements can still cause errors. A compatible cache handoff is optional, never assumed. The Drain and restart without abruptly dropping requests guide at /rust/instances/guides/instances.drain-restart provides the same confirmation, start and history checks.

Memory placement and engram offload

Open Memory and engram tiers at /rust/lexi-engine through the workflow page. Engram storage tier chooses automatic, host RAM, storage or a qualified device; slower tiers trade memory for latency. Expert cache budget, Expert placement and Prewarmed experts allocate RAM or device memory to reduce on-demand transfers. Resident memory budget and Host RAM reserve protect capacity. These are capability-gated: requesting an unsupported adapter or execution path fails rather than silently changing backend. Check the actual engine capabilities and saved-versus-running values after editing.

For measured tuning and automatic fit, see Panel-Optimizer.

Instance, profile and launch control reference

Each row below corresponds to a registered help marker (including the classic control when there is a matching classic selector). A feature's listed routes remain the authoritative paths for validation and persistence. The command-line column is populated only where a documented executable flag exists.

Control (feature ID)	Effect and cost	Native page	Routes	Command-line flag
Command-line parameter preview (instances.param-cmdline-import)	Parses launch flags into candidate parameters without writing settings or starting a service.	/rust/parameters	POST /api/params/from-cmdline	—
Installed template inventory (instances.template-inventory)	Shows template parsing, pinning and selection without changing saved settings.	/rust/templates	GET /api/templates	—
Choose chat template (instances.template-use)	Sets the selected saved template without restarting inference.	/rust/templates	POST /api/params	—
Save chat template (instances.template-save)	Validates a named template before saving it; existing content is guarded by the owner.	/rust/templates	POST /api/template	—
Extract model template (instances.template-extract)	Extracts a model's embedded template without selecting it.	/rust/templates	POST /api/template/extract	—
Download chat template (instances.template-download)	Downloads a template using verified transfer checks; no model is launched.	/rust/templates	POST /api/download	—
adaptive-p decay (instances.param-adaptive-decay)	How fast adaptive-p reacts. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
adaptive-p target (instances.param-adaptive-target)	adaptive-p: aims at a token probability and adapts; experimental upstream. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Model alias (instances.param-alias)	The model name clients see and send. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
API key (instances.param-api-key)	Required by clients when set; the raw port has no other login. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Compute backend (instances.param-backend)	Sets compute backend for this instance. Changing the runtime may require a restart and alter speed or compatibility.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Backend sampling (instances.param-backend-sampling)	Experimental upstream: samples on the GPU. Measure; a bad path can fail requests. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Logical batch (-b) (instances.param-batch)	Sets logical batch (-b) for this instance. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Cache idle slots (instances.param-cache-idle-slots)	Keeps idle conversations' caches so they resume without re-reading. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Prompt caching (instances.param-cache-prompt)	Sets prompt caching for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Prompt cache (host RAM) (instances.param-cache-ram)	Sets prompt cache (host ram) for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Cache reuse chunk (instances.param-cache-reuse)	Sets cache reuse chunk for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Checkpoint spacing (instances.param-checkpoint-min-step)	Wider spacing keeps fewer checkpoints: less memory, more recomputing. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Context shift (instances.param-context-shift)	On: long sessions go on by dropping the oldest text. Off: they stop at the limit. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
All MoE experts on CPU (instances.param-cpu-moe)	Every expert on the CPU: gets a big MoE model loading on a small card, slowly. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Context size (instances.param-ctx)	Sets context size for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Context checkpoints (instances.param-ctx-checkpoints)	More checkpoints let the server rewind instead of recomputing; each costs memory. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
DRY allowed length (instances.param-dry-allowed-length)	Repeats shorter than this are allowed; raise it for code. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
DRY base (instances.param-dry-base)	How fast DRY's penalty grows. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
DRY multiplier (instances.param-dry-multiplier)	DRY stops verbatim loops of whole phrases; off by default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
DRY window (instances.param-dry-penalty-last-n)	How far back DRY looks. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
dynatemp exponent (instances.param-dynatemp-exp)	Shape of the dynamic temperature. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
dynatemp range (instances.param-dynatemp-range)	Varies temperature with the model's certainty; off by default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Extra llama-server arguments (instances.param-extra-args)	Flags the form does not model, passed as typed (checked against this build's own list). Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Auto-fit to VRAM (instances.param-fit)	Sets auto-fit to vram for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Auto-fit min context (instances.param-fit-ctx)	The smallest context auto-fit may shrink to. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Auto-fit margin (instances.param-fit-target)	Memory the auto-fit leaves free. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Flash attention (instances.param-flash-attn)	On is faster and is what makes a quantized KV cache work. Off only to debug. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
frequency_penalty (instances.param-frequency-penalty)	Penalises frequent tokens. A default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
cuBLAS compute type (instances.param-ggml-cuda-cublas-compute-type)	Overrides matrix maths precision: speed against accuracy. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Device selection (instances.param-ggml-cuda-devices)	Sets device selection for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable kernel fusion (instances.param-ggml-cuda-disable-fusion)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable HIP graphs (instances.param-ggml-cuda-disable-graphs)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Unified memory (instances.param-ggml-cuda-enable-unified-memory)	Backs GPU memory with host RAM silently: runs, twenty times slower. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
HIP graph optimisation (instances.param-ggml-cuda-graph-opt)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable pinned host memory (instances.param-ggml-cuda-no-pinned)	Stops page-locked transfers: slower copies, less pinned RAM. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Register host memory (instances.param-ggml-cuda-register-host)	Sets register host memory for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Allow graphics queue (instances.param-ggml-vk-allow-graphics-queue)	Compute on the graphics queue when no compute queue is usable. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Async transfer queue (instances.param-ggml-vk-async-use-transfer-queue)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Debug markers (instances.param-ggml-vk-debug-markers)	Diagnostic logging; adds overhead, so switch it off when done. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable async (instances.param-ggml-vk-disable-async)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable bf16 (instances.param-ggml-vk-disable-bfloat16)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable coopmat (instances.param-ggml-vk-disable-coopmat)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable coopmat2 (instances.param-ggml-vk-disable-coopmat2)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable coopmat2 decode vector (instances.param-ggml-vk-disable-coopmat2-decode-vector)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable dot2 (instances.param-ggml-vk-disable-dot2)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable fp16 (instances.param-ggml-vk-disable-f16)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable kernel fusion (instances.param-ggml-vk-disable-fusion)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable graph optimise (instances.param-ggml-vk-disable-graph-optimize)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable host-visible VRAM (instances.param-ggml-vk-disable-host-visible-vidmem)	Changes how ReBAR memory is used. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable int dot (instances.param-ggml-vk-disable-integer-dot-product)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable MMVQ (instances.param-ggml-vk-disable-mmvq)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable multi-add (instances.param-ggml-vk-disable-multi-add)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Memory priority (instances.param-ggml-vk-enable-memory-priority)	Asks the driver to keep these allocations resident. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Disable sparse flash-attn (instances.param-ggml-vk-fa-sparse-disable)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Max allocation size (instances.param-ggml-vk-force-max-allocation-size)	Caps single allocations, a workaround for drivers that fail on large ones. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Force max buffer size (instances.param-ggml-vk-force-max-buffer-size)	Caps single buffers, a driver workaround. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Force MMVQ (instances.param-ggml-vk-force-mmvq)	Forces the quantized matrix-vector path; sometimes faster at one request. Measure. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Max nodes per submit (instances.param-ggml-vk-max-nodes-per-submit)	Sets max nodes per submit for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Memory logger (instances.param-ggml-vk-memory-logger)	Diagnostic logging; adds overhead, so switch it off when done. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Perf logger (instances.param-ggml-vk-perf-logger)	Diagnostic logging; adds overhead, so switch it off when done. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Perf logger concurrent (instances.param-ggml-vk-perf-logger-concurrent)	Diagnostic logging; adds overhead, so switch it off when done. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Perf logger frequency (instances.param-ggml-vk-perf-logger-frequency)	Diagnostic logging; adds overhead, so switch it off when done. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Pipeline statistics (instances.param-ggml-vk-pipeline-stats)	Diagnostic logging; adds overhead, so switch it off when done. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Prefer host memory (instances.param-ggml-vk-prefer-host-memory)	Puts the model in host RAM: the failure that collapses speed. Leave it unset. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Serialize submissions (instances.param-ggml-vk-serialize-submissions)	A driver switch for diagnosing one specific problem; most cost speed when changed. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Suballocation block (instances.param-ggml-vk-suballocation-block-size)	Allocation block size; affects fragmentation on a full card. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Sync logger (instances.param-ggml-vk-sync-logger)	Logs every sync event: fills the tmpfs log and RAM. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Visible devices (instances.param-ggml-vk-visible-devices)	Sets visible devices for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Bind address (instances.param-host)	Loopback keeps the service local; binding all interfaces exposes it to the network, so set an API key. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
SDMA copy engines (instances.param-hsa-enable-sdma)	Sets sdma copy engines for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Image max tokens (instances.param-image-max-tokens)	Caps tokens per image: saves context, may lose fine detail. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Image min tokens (instances.param-image-min-tokens)	More tokens per image: more detail read, more context used. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Tokens to keep (instances.param-keep)	Sets tokens to keep for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
KV cache on GPU (instances.param-kv-offload)	On keeps the KV cache in VRAM, which is much faster; off moves it to host RAM. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
KV cache quant (instances.param-kv-type)	Sets kv cache quant for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Unified KV buffer (instances.param-kv-unified)	One KV pool shared by every slot, so memory follows whichever requests are long. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
KV per slot (instances.param-kv-unified-per-slot)	Sets kv per slot for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Lazy tensor reads (instances.param-lazy-mode)	Sets lazy tensor reads for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Model load mode (instances.param-load-mode)	Sets model load mode for this instance. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Log verbosity (instances.param-log-verbosity)	How much the log says; debug levels fill memory. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
LoRA adapter (instances.param-lora)	An adapter on top of the model: changes how it behaves. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
LoRA scaled (instances.param-lora-scaled)	An adapter with a strength: file.gguf:0.5 applies it at half. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Main GPU index (instances.param-main-gpu)	Only for split mode none or row. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Max tool response chars (instances.param-max-tool-response-chars)	Shortens tool results in history to save context. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
min_p (instances.param-min-p)	Drops unlikely tokens relative to the best one. A default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Mirostat (instances.param-mirostat)	Replaces top-k, top-p and typical-p entirely; they are then ignored. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Mirostat tau (instances.param-mirostat-ent)	Mirostat's target surprise. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Mirostat eta (instances.param-mirostat-lr)	Mirostat's learning rate. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Projector file (instances.param-mmproj)	Sets projector file for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Vision projector device (instances.param-mmproj-device)	The vision projector on another device. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Projector on GPU (instances.param-mmproj-offload)	Projector on the CPU returns about 1.7 GB of VRAM; images are slower to read. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Model file (instances.param-model)	Which model runs. Changing it reloads the server. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Image encode batch (instances.param-mtmd-batch-max-tokens)	Image tokens encoded at once: smaller lowers the VRAM spike. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
GPU layers (instances.param-ngl)	Sets gpu layers for this instance. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Bypass host buffer (instances.param-no-host)	Changes where buffers live; can put the model in slow host memory. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
MoE layers on CPU (instances.param-n-cpu-moe)	The first N layers' experts on the CPU: raise until the model fits. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Max tokens per reply (instances.param-n-predict)	The longest reply the server allows. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Offload host ops (instances.param-op-offload)	Diagnostic: off moves work back to the CPU. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Tensor placement override (instances.param-override-tensors)	Sets tensor placement override for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Server slots (instances.param-parallel)	Conversations served at once; each gets its share of the context. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
API port (instances.param-port)	Where apps connect to this server. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
presence_penalty (instances.param-presence-penalty)	Pushes toward new topics. A default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Preserve thinking (instances.param-preserve-thinking)	Keeps every turn's thinking in context: better continuity, context fills fast. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Allow POST /props (instances.param-props)	Lets any client change server properties over HTTP. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Host RAM floor (instances.param-ram-floor-mb)	Stops the server before the host runs out of RAM and locks up. Keep it. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Reasoning mode (instances.param-reasoning)	Thinking on, off, or as the template decides. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Reasoning budget (instances.param-reasoning-budget)	Caps thinking tokens: answers start sooner, hard ones may be cut short. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Budget-exhausted message (instances.param-reasoning-budget-message)	Text inserted when the thinking budget runs out. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Reasoning effort (instances.param-reasoning-effort)	Higher thinks longer: better answers to hard questions, slower replies. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Reasoning format (instances.param-reasoning-format)	Where the thinking appears in the API response. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Weight repacking (instances.param-repack)	Only matters for CPU inference. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
repeat_last_n (instances.param-repeat-last-n)	How far back the repeat penalty looks. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
repeat_penalty (instances.param-repeat-penalty)	Sets repeat_penalty for this instance. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
RoPE freq base (instances.param-rope-freq-base)	Sets rope freq base for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
RoPE freq scale (instances.param-rope-freq-scale)	The inverse of RoPE scale: set one or the other, never both. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
RoPE scale (instances.param-rope-scale)	How far the context is stretched: 2 doubles it. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
RoPE scaling (instances.param-rope-scaling)	Sets rope scaling for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Sampler chain (instances.param-samplers)	The order samplers run in; the order changes the result. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
RNG seed (instances.param-seed)	Random per request. Pin it only to reproduce one output. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Sleep when idle (instances.param-sleep-idle-seconds)	Frees the GPU while idle; the first request afterwards waits for the reload. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Slots endpoint (instances.param-slots)	The slot view the panel reads. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft backend sampling (instances.param-spec-draft-backend-sampling)	Samples the draft on the GPU; on by default. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft model device (instances.param-spec-draft-device)	The draft on another card, so it stops competing with the model for memory. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft KV type (instances.param-spec-draft-kv-type)	KV cache type for the draft; a smaller type saves memory. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft model file (instances.param-spec-draft-model)	Sets draft model file for this instance. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft GPU layers (instances.param-spec-draft-ngl)	Draft layers on the GPU; set automatically. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-map-k4v min hits (instances.param-spec-ngram-map-k4v-min-hits)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-map-k4v draft M (instances.param-spec-ngram-map-k4v-size-m)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-map-k4v lookup N (instances.param-spec-ngram-map-k4v-size-n)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-map-k min hits (instances.param-spec-ngram-map-k-min-hits)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-map-k draft M (instances.param-spec-ngram-map-k-size-m)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-map-k lookup N (instances.param-spec-ngram-map-k-size-n)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-mod match len (instances.param-spec-ngram-mod-n-match)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-mod n-max (instances.param-spec-ngram-mod-n-max)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-mod n-min (instances.param-spec-ngram-mod-n-min)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-simple min hits (instances.param-spec-ngram-simple-min-hits)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-simple draft M (instances.param-spec-ngram-simple-size-m)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
ngram-simple lookup N (instances.param-spec-ngram-simple-size-n)	Tuning for the n-gram guess-ahead modes; only read when that spec type is chosen. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft depth (instances.param-spec-n-max)	How far each guess runs. Deeper helps only while guesses are accepted: measure. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft n-min (instances.param-spec-n-min)	Forces a minimum guess length even when unsure. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft p-min (instances.param-spec-p-min)	Confidence a guessed token needs before it is proposed. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Draft p-split (instances.param-spec-p-split)	Only for tree-style draft types. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Spec type (instances.param-spec-type)	Sets spec type for this instance. Drafting uses extra memory or compute and can lose speed if guesses fail.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Multi-GPU split mode (instances.param-split-mode)	How a model is split across cards: layer is the one for two different cards. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
SSE ping interval (instances.param-sse-ping-interval)	Keep-alive for streaming replies. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Full-size SWA cache (instances.param-swa-full)	Sliding-window models only: a full-size cache costs memory and helps cache reuse. More memory use can reduce capacity; changes take effect after restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Temperature (instances.param-temp)	Higher is more varied, lower more focused. A default: clients that send their own win. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Template checksum (instances.param-template-sha256)	Checks the saved template contents against a checksum; a mismatch can prevent loading or change replies.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Chat template (instances.param-template-src)	Sets chat template for this instance. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Tensor split (instances.param-tensor-split)	How much of the model each card takes. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
CPU threads (instances.param-threads)	CPU threads for work the GPU does not take: the number of physical cores. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
HTTP threads (instances.param-threads-http)	Threads for the web server, not the model. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Request timeout (instances.param-timeout)	Sets request timeout for this instance. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
top_k (instances.param-top-k)	Samples only among the k most likely tokens. A default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
top_n_sigma (instances.param-top-n-sigma)	Keeps tokens near the top score; off by default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
top_p (instances.param-top-p)	Sets top_p for this instance. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
typical_p (instances.param-typical-p)	Locally typical sampling; off by default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Physical batch (-ub) (instances.param-ubatch)	Sets physical batch (-ub) for this instance. Higher throughput settings may use more memory or harm latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Web UI defaults file (instances.param-ui-config-file)	The built-in web UI's starting settings. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Web UI MCP proxy (instances.param-ui-mcp-proxy)	Relays the web UI's tool traffic; experimental upstream. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Enable vision (instances.param-use-mmproj)	Off frees VRAM but the model can no longer read images. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Built-in web UI (instances.param-webui)	Sets built-in web ui for this instance. Changes may interrupt connections or require a restart.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
XTC probability (instances.param-xtc-probability)	Drops the top choices for variety, at the cost of accuracy; off by default. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
XTC threshold (instances.param-xtc-threshold)	Which top choices XTC may drop. This can change response quality, determinism, or latency.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
YaRN attn factor (instances.param-yarn-attn-factor)	A YaRN fine-tuning dial: automatic (-1) is right. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
YaRN beta fast (instances.param-yarn-beta-fast)	A YaRN fine-tuning dial: automatic (-1) is right. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
YaRN beta slow (instances.param-yarn-beta-slow)	A YaRN fine-tuning dial: automatic (-1) is right. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
YaRN ext factor (instances.param-yarn-ext-factor)	A YaRN fine-tuning dial: automatic (-1) is right. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
YaRN original ctx (instances.param-yarn-orig-ctx)	Sets yarn original ctx for this instance. Changes can affect memory use, output quality, speed, or availability.	/rust/parameters	GET /api/param-meta, GET /api/params, POST /api/params	—
Instance inventory (instances.inventory)	Lists saved instances and live state; refreshing makes a read without restarting anything.	/rust/instances/workflows	GET /api/instances	—
Create instance (instances.create)	Creates a separate instance from selected devices; a new instance needs disk space and a free listener.	/rust/instances/workflows	POST /api/instance/create	—
Device selection (instances.device)	Selects any valid GPU subset or CPU-only; moving devices can require a stop and changes memory capacity.	/rust/instances/workflows	POST /api/instance/update	—
Per-instance engine (instances.engine)	Chooses the runtime for a new instance; available acceleration and required libraries depend on the engine.	/rust/instances/workflows	POST /api/instance/create	—
Auto-fit preview (instances.autofit)	Estimates fit from installed devices without applying settings; estimates may differ from actual loading.	/rust/instances/workflows	POST /api/instance/create	—
Start instance (instances.start)	Starts the selected service; model loading consumes device or host memory and can fail admission.	/rust/instances/workflows	POST /api/start	—
Stop instance (instances.stop)	Stops serving the selected instance; in-flight requests can be interrupted.	/rust/instances/workflows	POST /api/stop	—
Immediate restart (instances.restart)	Restarts the service without a drain; active requests can be interrupted.	/rust/instances/workflows	POST /api/restart	—
Safe restart (instances.safe-restart)	Holds new traffic, drains active requests and verifies the replacement; requests may wait or fail on timeout.	/rust/instances/workflows	POST /api/restarts/start	—
Request drain (instances.drain)	Allows current requests time to finish before a restart; excess traffic may be refused during the hold.	/rust/instances/workflows	POST /api/restarts/settings	—
Gateway hold (instances.restart-hold)	Sets how long new requests are held during restart; longer holds can increase client waiting.	/rust/instances/workflows	POST /api/restarts/settings	—
Restart verification (instances.restart-canary)	Checks a replacement before releasing traffic; verification adds time but reduces bad activations.	/rust/instances/workflows	POST /api/restarts/settings	—
Cache handoff (instances.restart-kv)	Attempts compatible request-cache restoration during restart; snapshots use time and storage.	/rust/instances/workflows	POST /api/restarts/settings	—
Restart report (instances.restart-report)	Shows restart outcome and rollback evidence; it makes no configuration change.	/rust/instances/workflows	GET /api/restarts/report	—
Instance profiles (instances.profile-list)	Shows reusable saved settings; loading them can change device and memory requirements.	/rust/instances/workflows	GET /api/instance-profiles	—
Save instance profile (instances.profile-save)	Copies a selected instance configuration to a named profile; saved copies consume storage.	/rust/instances/workflows	POST /api/instance-profile/save	—
Apply instance profile (instances.profile-apply)	Writes selected settings and tiers, with rollback protection; live changes may require restart.	/rust/instances/workflows	POST /api/instance-profile/apply	—
Copy instance profile (instances.profile-copy)	Makes another named copy of saved settings; it consumes storage and does not restart serving.	/rust/instances/workflows	POST /api/instance-profile/copy	—
Rename instance profile (instances.profile-rename)	Changes a profile label; linked procedures may need to use the new name.	/rust/instances/workflows	POST /api/instance-profile/rename	—
Delete instance profile (instances.profile-delete)	Removes reusable settings; running instances keep their own configuration.	/rust/instances/workflows	POST /api/instance-profile/delete	—
Fallback tiers (instances.tier-list)	Shows normal, safe, and minimal fallback settings; smaller tiers can reduce context or speed.	/rust/parameters	GET /api/tiers	—
Load tier (instances.tier-load)	Loads a saved fallback into the editor without restarting; nothing changes until saved.	/rust/parameters	POST /api/tier/load	—
Save tier (instances.tier-save)	Saves a fallback parameter set; replacing a tier changes future recovery behavior.	/rust/parameters	POST /api/tier/save	—
Copy tier (instances.tier-copy)	Copies saved parameters to a fallback tier; overwriting a tier removes its old values.	/rust/parameters	POST /api/tier/copy	—
Delete tier (instances.tier-delete)	Removes a fallback profile; a failed launch may lose this recovery choice.	/rust/parameters	POST /api/tier/delete	—
Reset launch failures (instances.tier-reset)	Resets the failure counter; this can allow a higher-memory tier to be tried again.	/rust/parameters	POST /api/tier/reset-fails	—