werk1112

Complete CLI command and option index

Build scope: canonical Linux, --no-default-features, no optional backend features. Coverage is limited to this parser/schema profile, not every platform or feature build. See conditional build options for extensions.

Generated from werk 1.10.2. See CLI semantics, media input and override rules, and the canonical parameter schema.

Coverage of this profile: 31 top-level commands, 100 help nodes (including the root), and 455 distinct long option names (excluding --help and --version). Compatibility aliases are listed in CLI conventions.

<VALUE> is required; [VALUE] is optional; ... permits multiple values. Usage lines show required arguments/options. Defaults and accepted enum values below come from the parser. An absent default means no parser value is inserted; runtime resolution may still supply one. The built-in werk help COMMAND … and werk COMMAND … --help display command help.

Command index

• werk support
• werk gpus
• werk deployment-plan
• werk top
• werk serve
• werk run
• werk chat
• werk image
• werk video
• werk audio
• werk estimate
• werk bench
• werk doctor
• werk backend
• werk updates
• werk components
• werk update
• werk artifacts
• werk auth
• werk temp
• werk cache
• werk runtime
• werk import
• werk convert
• werk pull
• werk setup
• werk remove
• werk list
• werk parameters
• werk inspect
• werk select-file
• werk support list
• werk support register
• werk support register-action
• werk support image
• werk support speech
• werk support action
• werk image generate
• werk image edit
• werk image upscale
• werk video generate
• werk video animate
• werk video transform
• werk video upscale
• werk audio generate
• werk audio transcribe
• werk audio translate
• werk audio detect
• werk audio analyze
• werk audio transform
• werk audio embed
• werk doctor perf
• werk backend update
• werk backend updates
• werk backend install
• werk backend list
• werk backend doctor
• werk artifacts vision
• werk artifacts build
• werk artifacts list
• werk artifacts rebuild
• werk auth huggingface
• werk auth api-key
• werk temp list
• werk temp purge
• werk temp path
• werk cache list
• werk cache purge
• werk runtime info
• werk runtime capabilities
• werk runtime memory
• werk runtime states
• werk runtime state
• werk runtime prune
• werk audio generate speech
• werk audio generate music
• werk audio generate sound
• werk audio detect event
• werk audio detect voice
• werk audio detect speaker
• werk audio detect language
• werk audio detect emotion
• werk audio analyze caption
• werk audio analyze diarize
• werk audio analyze classify
• werk audio analyze understand
• werk audio transform voice
• werk audio transform separate
• werk audio transform enhance
• werk audio transform edit
• werk auth huggingface login
• werk auth huggingface status
• werk auth huggingface logout
• werk auth api-key generate
• werk runtime state pin
• werk runtime state unpin
• werk runtime state promote
• werk runtime state demote
• werk runtime state evict

werk

Werk1112 imports local or Hugging Face models, routes inference across installed runtimes, serves OpenAI-compatible and Anthropic Messages API subsets, and provides a live terminal dashboard with werk top.

werk [OPTIONS] [COMMAND]

Subcommands: werk support, werk gpus, werk deployment-plan, werk top, werk serve, werk run, werk chat, werk image, werk video, werk audio, werk estimate, werk bench, werk doctor, werk backend, werk updates, werk components, werk update, werk artifacts, werk auth, werk temp, werk cache, werk runtime, werk import, werk convert, werk pull, werk setup, werk remove, werk list, werk parameters, werk inspect, werk select-file.

Global options

Option	Meaning / parser default
--log-level <LOG_LEVEL>	Operational log threshold. –verbose/–verbose-pure default to debug [env: WERK_LOG_LEVEL] [possible values: error, warn, info, debug, trace, off]
--log-format <LOG_FORMAT>	stderr rendering; auto uses the terminal palette and plain text in pipes [env: WERK_LOG_FORMAT] [default: auto] [possible values: auto, text, json]
--log-file <LOG_FILE>	Append JSONL events to this file, independently of terminal rendering [env: WERK_LOG_FILE]
--log-max-size-mb <LOG_MAX_SIZE_MB>	Rotate the JSONL file after this many MiB (one writer per path) [env: WERK_LOG_MAX_SIZE_MB] [default: 10]
--log-retention <LOG_RETENTION>	Number of rotated files to retain, named PATH.1 through PATH.N [env: WERK_LOG_RETENTION] [default: 5]
--deployments <DEPLOYMENTS>	Server-owned deployment profiles JSON (serve only) [env: WERK_DEPLOYMENTS]
--model-home <MODEL_HOME>	Model store directory; defaults to WERK_HOME, XDG_DATA_HOME/werk1112, or ~/.local/share/werk1112 [env: WERK_HOME]
--device <DEVICE>	Candle-only device override for this command: auto, cpu, cuda, or metal [possible values: auto, cpu, cuda, metal]
--backend <BACKEND>	Backend for this process: auto, cpu, cuda, rocm, vulkan, metal, mlx, omlx, onnx, transformers, vllm, candle, llama-highlevel, or llama-legacy [default: auto] [possible values: auto, candle, cpu, cuda, llama-highlevel, llama-legacy, metal, mlx, omlx, onnx, rocm, transformers, vllm, vulkan]
--auto-install-backends	Approve installation of missing model backends and feature components (including vision projectors)
--no-auto-install-backends	Disable automatic managed runtime provisioning
--parallel <PARALLEL>	llama-server active sequence slots; default 1; each slot receives –ctx-size tokens [env: WERK_LLAMA_PARALLEL]
--fp4-kernel <FP4_KERNEL>	NVFP4 CUDA kernel policy; default auto prefers native Blackwell, then Marlin, then GGML [env: WERK_FP4_KERNEL] [possible values: auto, native, marlin, ggml]
--ctx-size <CTX_SIZE>	llama.cpp context size per sequence; 0 uses the model default [env: WERK_LLAMA_CTX]
--batch-size <BATCH_SIZE>	llama.cpp logical prompt batch size [env: WERK_LLAMA_BATCH]
--ubatch-size <UBATCH_SIZE>	llama.cpp physical compute micro-batch size [env: WERK_LLAMA_UBATCH]
--gpu-layers <GPU_LAYERS>	llama.cpp GPU layers; high values mean all layers [env: WERK_LLAMA_GPU_LAYERS]
--main-gpu <MAIN_GPU>	llama.cpp main GPU index [env: WERK_LLAMA_MAIN_GPU]
--kv-cache-type <KV_CACHE_TYPE>	llama.cpp KV cache type: f16, f32, or q8-0 [env: WERK_LLAMA_KV_CACHE_TYPE] [possible values: f16, f32, q8-0]
--flash-attn [<FLASH_ATTN>]	Request llama.cpp flash attention when the native runtime exposes it [env: WERK_LLAMA_FLASH_ATTN] [possible values: true, false]
--kv-offload [<KV_OFFLOAD>]	Control llama.cpp K/Q/V and KV-cache GPU offload [env: WERK_LLAMA_KV_OFFLOAD] [possible values: true, false]
--warmup-tokens <WARMUP_TOKENS>	Synthetic warmup tokens; 0 disables prewarm. llama-server defaults to off and treats positive values as enabling native warmup [env: WERK_LLAMA_WARMUP_TOKENS]
--threads <THREADS>	llama.cpp generation CPU helper threads [env: WERK_LLAMA_THREADS]
--threads-batch <THREADS_BATCH>	llama.cpp prompt-eval CPU helper threads [env: WERK_LLAMA_THREADS_BATCH]
-h, --help	Print help (see a summary with ‘-h’)
-V, --version	Print version

Global options and --help are inherited by subcommands unless a local declaration overrides them. Their backend scope is described in Global options.

werk support

Official model support catalog, hosted inference and action prediction

werk support [OPTIONS] <COMMAND>

Subcommands: werk support list, werk support register, werk support register-action, werk support image, werk support speech, werk support action.

Also accepts the global options.

werk gpus

Print physical GPU inventory and inherited visibility as JSON

werk gpus [OPTIONS]

Also accepts the global options.

werk deployment-plan

Resolve deployment profiles without loading weights or starting workers

werk deployment-plan [OPTIONS] <CONFIG>

Arguments:

<CONFIG>

Option	Meaning / parser default
--inventory <INVENTORY>	Read an inventory JSON fixture instead of probing hardware (dry-run only)

Also accepts the global options.

werk top

Live Werk dashboard with inference, memory and cache telemetry

werk top [OPTIONS]

Option	Meaning / parser default
--url <URL>	Server root URL without /v1, credentials, query or fragment; HTTP(S). [default: http://127.0.0.1:11434]
--api-key <API_KEY>	Bearer credential; local servers also support discovery from the default key file. [env: WERK_API_KEY]
--interval-ms <INTERVAL_MS>	Refresh interval in milliseconds; range 500–60000. [default: 2000]
--once	Print one snapshot and exit, without entering terminal mode
--json	Print the snapshot as JSON; requires –once.
--no-animation	Disable decorative dashboard animation.
--demo	Preview the dashboard with explicitly simulated data; no server connection

Also accepts the global options.

werk serve

Start the OpenAI-compatible HTTP server

werk serve [OPTIONS]

Option	Meaning / parser default
--host <HOST>	Address to bind [default: 127.0.0.1]
--port <PORT>	Port to bind [default: 11434]
--model <MODEL>	Default model for chat or text-analysis API requests that omit model
--preload <MODEL>...	Load, warm and retain local models in order; hosted models prepare clients without paid inference; repeatable, startup exits on failure
--image-model <IMAGE_MODEL>	Default image model for compatible clients that omit model or use an OpenAI image alias
--api-key <API_KEY>	Require OpenAI-style Authorization: Bearer for /v1 requests [env: WERK_API_KEY]
--api-keys <PATH>	Load OpenAI-style bearer keys from an API keys TOML file [env: WERK_API_KEYS]
--allow-unauthenticated	Disable API key auth for local development
--cors-origin <ORIGIN>	Allow an exact browser origin; repeat for multiple origins (wildcard and null are rejected)
--verbose	Debug logs and detailed terminal statistics
--verbose-pure	Raw JSONL debug logs, without terminal presentation [aliases: –verbose-lite]
--persistence	Enable persistence defaults for /werk/v1/prefill, oMLX native short-prefix caching during the worker lifetime (auto/disk modes), and automatic prefix caching for local Werk-managed vLLM; remote vLLM remains externally managed
--persistence-mode <PERSISTENCE_MODE>	Default prefill persistence mode when policy is omitted; implies –persistence [possible values: ephemeral, memory, disk, auto]
--persistence-reuse <PERSISTENCE_REUSE>	Default prefill reuse policy when policy is omitted; disabled also disables the managed local-vLLM prefix-cache default and oMLX native short-prefix caching; implies –persistence [possible values: disabled, prefer, required]
--persistence-ttl-seconds <PERSISTENCE_TTL_SECONDS>	Default prefill state TTL in seconds when policy is omitted; implies –persistence
--persistence-pin	Pin prefill state by default when policy is omitted; implies –persistence

Also accepts the global options.

werk run

Run text, vision, tools, image, audio, or video inference with an installed model

werk run [OPTIONS] <MODEL> [PROMPT]...

Arguments:

<MODEL>      Installed model id
  [PROMPT]...  Prompt text

Option	Meaning / parser default
--max-tokens <MAX_TOKENS>	Maximum generated tokens (default: 256, or request JSON value)
--temperature <TEMPERATURE>	Sampling temperature
--top-p <TOP_P>	Nucleus sampling top-p
--seed <SEED>	RNG seed
--chat-template <CHAT_TEMPLATE>	Override chat templating: model, generic, phi3, llama3, gemma, chatml, qwen-chatml, or none [possible values: model, generic, phi3, llama3, gemma, chatml, qwen-chatml, none]
--image <PATH_OR_URL>	Attach an image for VLM-capable backends; may be repeated
--verbose	Print Ollama-style timing and throughput stats
--debug	Print backend internals and resolved runtime details
--server <URL>	Send text/vision/tool requests to an existing werk serve (http://HOST:PORT); configure the worker on serve
--api-key <API_KEY>	Bearer key for –server [env: WERK_API_KEY]
--request <PATH>	Read an OpenAI chat request or canonical media InferenceRequest from JSON; - reads stdin
--task <TASK>	Inference task; inferred when the model has an unambiguous default
--input <MODALITY[:ROLE]=PATH_OR_URL>	Media input, e.g. audio=clip.wav or image:mask_image=mask.png; repeatable
--set <PATH=VALUE>	Canonical media parameter, e.g. image.steps=20 or routing.precision=float16; repeatable
--output <PATH>	Publish media output to a file or directory
--json	Print structured results; with text streaming, emit NDJSON events
--stream	Stream generated text; with –json, also stream tool-call deltas
--stream-granularity <STREAM_GRANULARITY>	Text stream granularity: token or chunk; implies –stream [possible values: token, chunk]
--no-history	Use only this invocation’s messages; conflicts with conversation persistence
--persistence	Save and resume completed chat turns across restarts for every backend; native KV reuse depends on the backend
--session <NAME>	Persistent conversation name for this model (default: default); implies –persistence
--persistence-reuse <PERSISTENCE_REUSE>	Conversation reuse: prefer resumes if present, required needs a saved conversation, disabled starts fresh; implies –persistence [possible values: disabled, prefer, required]
--persistence-mode <PERSISTENCE_MODE>	Conversation storage: disk/auto survives exit; memory/ephemeral lasts only for this process; implies –persistence [possible values: ephemeral, memory, disk, auto]
--persistence-ttl-seconds <PERSISTENCE_TTL_SECONDS>	Expire saved conversation after this many seconds; implies –persistence

Also accepts the global options.

werk chat

Start an interactive terminal chat with an installed model

werk chat [OPTIONS] <MODEL>

Arguments:

<MODEL>  Installed model id

Option	Meaning / parser default
--max-tokens <MAX_TOKENS>	Maximum generated tokens per turn [default: 256]
--temperature <TEMPERATURE>	Sampling temperature
--top-p <TOP_P>	Nucleus sampling top-p
--seed <SEED>	RNG seed
--chat-template <CHAT_TEMPLATE>	Override chat templating: model, generic, phi3, llama3, gemma, chatml, qwen-chatml, or none [possible values: model, generic, phi3, llama3, gemma, chatml, qwen-chatml, none]
--no-history	Do not include previous chat turns in the next prompt
--image <PATH_OR_URL>	Attach images to each turn for VLM-capable backends; may be repeated
--stream-granularity <STREAM_GRANULARITY>	How terminal chat streams text: token or chunk [default: token] [possible values: token, chunk]
--verbose	Print Ollama-style timing and throughput stats
--debug	Print backend internals and resolved runtime details
--persistence	Save and resume completed chat turns across restarts for every backend; native KV reuse depends on the backend
--session <NAME>	Persistent conversation name for this model (default: default); implies –persistence
--persistence-reuse <PERSISTENCE_REUSE>	Conversation reuse: prefer resumes if present, required needs a saved conversation, disabled starts fresh; implies –persistence [possible values: disabled, prefer, required]
--persistence-mode <PERSISTENCE_MODE>	Conversation storage: disk/auto survives exit; memory/ephemeral lasts only for this process; implies –persistence [possible values: ephemeral, memory, disk, auto]
--persistence-ttl-seconds <PERSISTENCE_TTL_SECONDS>	Expire saved conversation after this many seconds; implies –persistence

Also accepts the global options.

werk estimate

Estimate whether a local or Hugging Face model is likely to fit in memory

werk estimate [OPTIONS] <MODEL>

Arguments:

<MODEL>  Installed model id or Hugging Face repo id

Option	Meaning / parser default
--task <TASK>	Estimate an inference task using the canonical media request pipeline
--width <WIDTH>	Requested output width
--height <HEIGHT>	Requested output height
--frames <FRAMES>	Requested video frame count
--duration <SECONDS>	Requested output duration
--batch-size <BATCH_SIZE>	Requested batch size
--sample-rate <SAMPLE_RATE>	Requested audio sample rate
--channels <CHANNELS>	Requested audio channel count
--steps <STEPS>	Requested diffusion or sampling step count
--file <FILE>	For remote Hugging Face estimates, select a repository file; matching GGUF shards are included automatically
--json	Print machine-readable estimate JSON
--verbose	Print weight-file accounting details

Also accepts the global options.

werk bench

Benchmark an installed model backend

werk bench [OPTIONS] --prompt <PROMPT> <MODEL>

Arguments:

<MODEL>  Installed model id

Option	Meaning / parser default
--prompt <PROMPT>	Prompt text for the benchmark
--max-tokens <MAX_TOKENS>	Maximum generated tokens per run [default: 256]
--runs <RUNS>	Measured runs [default: 5]
--warmups <WARMUPS>	Warmup runs before measurement [default: 1]
--temperature <TEMPERATURE>	Benchmark sampling temperature; 0 uses greedy decoding [default: 0]
--top-p <TOP_P>	Benchmark nucleus sampling top-p
--seed <SEED>	Benchmark RNG seed [default: 42]
--compare <COMPARE>	Also benchmark another backend family [default: none] [possible values: none, legacy]
--print-native-info	Print resolved llama.cpp runtime settings
--json	Print machine-readable benchmark JSON
--include-output	Include generated text in benchmark JSON for quality review
--debug	Print backend internals during benchmark runs

Also accepts the global options.

werk doctor

Inspect Werk runtime diagnostics

werk doctor [OPTIONS] [COMMAND]

Subcommands: werk doctor perf.

Option	Meaning / parser default
--task <TASK>	Limit media diagnostics to an inference task
--runtime <RUNTIME>	Limit diagnostics to a runtime name
--model <MODEL>	Inspect routing for an installed model

Also accepts the global options.

werk backend

Manage local runtime backends

werk backend [OPTIONS] <COMMAND>

Subcommands: werk backend update, werk backend updates, werk backend install, werk backend list, werk backend doctor.

Also accepts the global options.

werk updates

Scan installed backends, models and vision projectors for updates without changing them

werk updates [OPTIONS]

Option	Meaning / parser default
--json	Print the complete report as JSON, including per-component errors
--offline	Inventory local components without contacting upstream registries
--scope <SCOPE>	Components to scan [default: all] [possible values: all, backends, artifacts]

Also accepts the global options.

werk components

Browse, install and update host-compatible components in a terminal UI

werk components [OPTIONS]

Option	Meaning / parser default
--list	Print the component inventory instead of opening the terminal UI
--json	Print host compatibility, installed versions, status, requirements and planned actions as JSON; requires –list and executes no actions.
--offline	Use local metadata only; installation and update actions stay disabled

Also accepts the global options.

werk update

Update a managed component or all installed compatible components, including dependencies

werk update [OPTIONS] [COMPONENT]

Arguments:

[COMPONENT]  Backend target or model:MODEL_ID; omit with --all

Option	Meaning / parser default
--all	Update all installed, host-compatible managed backends and changed models
--include-missing	Also set up missing dependencies required by installed models
-y, --yes	Approve the displayed plan without prompting
--dry-run	Display the concrete plan without changing components

Also accepts the global options.

werk artifacts

Manage optimized runtime artifacts for installed models

werk artifacts [OPTIONS] <COMMAND>

Subcommands: werk artifacts vision, werk artifacts build, werk artifacts list, werk artifacts rebuild.

Also accepts the global options.

werk auth

Manage external service authentication

werk auth [OPTIONS] <COMMAND>

Subcommands: werk auth huggingface, werk auth api-key.

Also accepts the global options.

werk temp

Manage temporary files in the model store

werk temp [OPTIONS] <COMMAND>

Subcommands: werk temp list, werk temp purge, werk temp path.

Also accepts the global options.

werk cache

List and remove local persistence caches

werk cache [OPTIONS] <COMMAND>

Subcommands: werk cache list, werk cache purge.

Also accepts the global options.

werk runtime

Inspect and control a running Werk runtime

werk runtime [OPTIONS] <COMMAND>

Subcommands: werk runtime info, werk runtime capabilities, werk runtime memory, werk runtime states, werk runtime state, werk runtime prune.

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode
--url <URL>	Werk server base URL [default: http://127.0.0.1:11434]
--api-key <API_KEY>	Bearer key for the Werk server [env: WERK_API_KEY]
--timeout-seconds <TIMEOUT_SECONDS>	Timeout in seconds for one runtime-control request [default: 30]

Also accepts the global options.

werk import

Import local models, copying their files or binding external paths

werk import [OPTIONS] <PATH>

Arguments:

<PATH>  Model file or directory, or a collection directory with --all

Option	Meaning / parser default
--name <NAME>	Installed model id; required unless –all is used
--all	Import each model directly inside a collection directory, preserving existing Werk model ids
--link	Register the existing model in place without copying its files
--vision	Set up image input for this backend/model; does not prove image inference
--text-only	Skip optional image components and keep this setup text-only
--offline	Do not contact repositories or install packages (existing-asset setup only)

Also accepts the global options.

werk convert

Convert an installed NVFP4 Hugging Face model to a separate GGUF model

werk convert [OPTIONS] --name <NAME> <MODEL>

Arguments:

<MODEL>  Installed source model id

Option	Meaning / parser default
--to <TO>	Destination model format. [default: gguf] [possible values: gguf]
--name <NAME>	New model id; existing models are never overwritten
--converter <CONVERTER>	Path to llama.cpp convert_hf_to_gguf.py
--python <PYTHON>	Python interpreter with converter dependencies installed
--verbose	Show converter diagnostics.

Also accepts the global options.

werk pull

Pull a Hugging Face repository into the managed model store

werk pull [OPTIONS] <REPO>

Arguments:

<REPO>  Hugging Face repo id, for example org/model

Option	Meaning / parser default
--name <NAME>	Installed model id; defaults to the repo id
--file <FILE>	Download a repository file, for example model.Q4_K_M.gguf; matching GGUF shards are included automatically
--vision	Set up image input for this backend/model; does not prove image inference
--text-only	Skip optional image components and keep this setup text-only
--offline	Do not contact repositories or install packages (existing-asset setup only)
--projector <PROJECTOR>	Select an exact model-specific projector file from this repository

Also accepts the global options.

werk setup

Check all model features and offer to install missing backends and components

werk setup [OPTIONS] <MODEL>

Arguments:

<MODEL>

Option	Meaning / parser default
--vision	Set up image input for this backend/model; does not prove image inference
--text-only	Skip optional image components and keep this setup text-only
--offline	Do not contact repositories or install packages (existing-asset setup only)

Also accepts the global options.

werk remove

Remove an installed model from the managed model store

werk remove [OPTIONS] <ID>

Arguments:

<ID>  Installed model id

Also accepts the global options.

werk list

List installed models and their local storage paths

werk list [OPTIONS]

Option	Meaning / parser default
--task <TASK>	Filter by supported task
--input-modality <INPUT>	Filter by input modality
--output-modality <OUTPUT>	Filter by output modality
--family <FAMILY>	Filter by model family
--layout <LAYOUT>	Filter by repository layout
--json	Print manifests as JSON

Also accepts the global options.

werk parameters

Describe task parameters, defaults, and resolution sources

werk parameters [OPTIONS] [MODEL]

Arguments:

[MODEL]  Installed model id

Option	Meaning / parser default
--task <TASK>	Inference task whose parameter contract should be shown
--json	Print machine-readable parameter information
--example	Print a canonical request example
--sources	Resolve and include parameter provenance

Also accepts the global options.

werk inspect

Print a model manifest as JSON

werk inspect [OPTIONS] <ID>

Arguments:

<ID>  Installed model id

Also accepts the global options.

werk select-file

Select which tracked model file an installed manifest uses

werk select-file [OPTIONS] <ID> <FILE>

Arguments:

<ID>    Installed model id
  <FILE>  Model file path, for example model.Q4_K_M.gguf or files/model.Q4_K_M.gguf

Also accepts the global options.

werk support list

List upstream identity, implemented capabilities and honest verification status

werk support list [OPTIONS]

Also accepts the global options.

werk support register

Register an official hosted model; credentials remain in environment variables

werk support register [OPTIONS] <MODEL>

Arguments:

<MODEL>

Option	Meaning / parser default
--name <NAME>	Optional local model ID; defaults to the official model ID.

Also accepts the global options.

werk support register-action

Register a downloaded DROID policy package without copying weights

werk support register-action [OPTIONS] --checkpoint <CHECKPOINT> <NAME>

Arguments:

<NAME>

Option	Meaning / parser default
--checkpoint <CHECKPOINT>	Required local DROID package directory containing config.json and model.safetensors.

Also accepts the global options.

werk support image

Generate/edit with FLUX 3 Image; request uses the documented BFL fields

werk support image [OPTIONS] --request <REQUEST> --output <OUTPUT>

Option	Meaning / parser default
--request <REQUEST>	Required JSON file with the BFL image request; see the model support guide linked from CLI semantics.
--output <OUTPUT>	Required new output image file; an existing path is rejected.

Also accepts the global options.

werk support speech

Generate Eleven v4 dialogue audio; request requires text/voice_id pairs

werk support speech [OPTIONS] --request <REQUEST> --output <OUTPUT>

Option	Meaning / parser default
--request <REQUEST>	Required JSON file with text/voice_id dialogue inputs; see the model support guide linked from CLI semantics.
--output <OUTPUT>	Required new output audio file; an existing path is rejected.

Also accepts the global options.

werk support action

Predict DROID actions from synchronized base64 camera images and joint state

werk support action [OPTIONS] --request <REQUEST> --output <OUTPUT>

Option	Meaning / parser default
--request <REQUEST>	Required JSON file with model, synchronized camera images and joint state; see the model support guide linked from CLI semantics.
--output <OUTPUT>	Required new output JSON file; an existing path is rejected.

Also accepts the global options.

werk doctor perf

Print llama.cpp performance/runtime diagnostics for an installed model

werk doctor perf [OPTIONS] <MODEL>

Arguments:

<MODEL>  Installed model id

Also accepts the global options.

werk backend update

Update a managed backend and its compatible dependencies

werk backend update [OPTIONS] <TARGET>

Arguments:

<TARGET>
          Possible values:
          - llama-cuda
          - llama-cuda-nvfp4:   Pinned CUDA runtime with native NVFP4 and standalone Marlin kernels
          - llama-cuda-offload: Experimental pinned CUDA runtime with MoE caching and lazy PLE reads
          - llama-rocm
          - llama-vulkan
          - llama-metal
          - llama-cpu
          - onnx-cuda
          - onnx-rocm
          - onnx-cpu
          - vllm
          - omlx:               Install the managed oMLX CLI on Apple Silicon (explicit installation only)
          - mlx:                Install official MLX-LM on Apple Silicon
          - mlx-vlm:            Install MLX-VLM on Apple Silicon
          - transformers:       Install Hugging Face Transformers and PyTorch
          - media:              Install Diffusers, Transformers and media codecs
          - qwen-tts
          - text-analysis

Option	Meaning / parser default
--vision	Set up image input for this backend/model; does not prove image inference
--text-only	Skip optional image components and keep this setup text-only
--offline	Do not contact repositories or install packages (existing-asset setup only)

Also accepts the global options.

werk backend updates

Scan installed managed backends for updates without installing them

werk backend updates [OPTIONS]

Option	Meaning / parser default
--json	Print the complete report as JSON, including per-component errors
--offline	Inventory local components without contacting upstream registries

Also accepts the global options.

werk backend install

Install a managed runtime backend

werk backend install [OPTIONS] <BACKEND>

Arguments:

<BACKEND>
          Backend to install, for example llama-cuda, onnx-cuda, vllm, or qwen-tts

          Possible values:
          - llama-cuda
          - llama-cuda-nvfp4:   Pinned CUDA runtime with native NVFP4 and standalone Marlin kernels
          - llama-cuda-offload: Experimental pinned CUDA runtime with MoE caching and lazy PLE reads
          - llama-rocm
          - llama-vulkan
          - llama-metal
          - llama-cpu
          - onnx-cuda
          - onnx-rocm
          - onnx-cpu
          - vllm
          - omlx:               Install the managed oMLX CLI on Apple Silicon (explicit installation only)
          - mlx:                Install official MLX-LM on Apple Silicon
          - mlx-vlm:            Install MLX-VLM on Apple Silicon
          - transformers:       Install Hugging Face Transformers and PyTorch
          - media:              Install Diffusers, Transformers and media codecs
          - qwen-tts
          - text-analysis

Option	Meaning / parser default
--vision	Set up image input for this backend/model; does not prove image inference
--text-only	Skip optional image components and keep this setup text-only
--offline	Do not contact repositories or install packages (existing-asset setup only)

Also accepts the global options.

werk backend list

List discovered runtime backends

werk backend list [OPTIONS]

Also accepts the global options.

werk backend doctor

Check runtime discovery and managed backend prerequisites

werk backend doctor [OPTIONS]

Option	Meaning / parser default
--debug	Print detailed backend discovery paths and rejection reasons

Also accepts the global options.

werk artifacts vision

Opt in to image input using existing model assets; never downloads the main model

werk artifacts vision [OPTIONS] <MODEL>

Arguments:

<MODEL>

Option	Meaning / parser default
--vision	Set up image input for this backend/model; does not prove image inference
--text-only	Skip optional image components and keep this setup text-only
--offline	Do not contact repositories or install packages (existing-asset setup only)
--projector <PROJECTOR>	Explicitly associate a matching local GGUF projector with this model
--download	Download only a projector from this model’s recorded source revision
--file <FILE>	Exact projector filename at the recorded repository revision

Also accepts the global options.

werk artifacts build

Build missing optimized artifacts for an installed model

werk artifacts build [OPTIONS] <MODEL>

Arguments:

<MODEL>  Installed model id

Also accepts the global options.

werk artifacts list

List optimized artifacts for an installed model

werk artifacts list [OPTIONS] <MODEL>

Arguments:

<MODEL>  Installed model id

Also accepts the global options.

werk artifacts rebuild

Rebuild optimized artifacts for an installed model

werk artifacts rebuild [OPTIONS] <MODEL>

Arguments:

<MODEL>  Installed model id

Also accepts the global options.

werk auth huggingface

Manage Hugging Face authentication

werk auth huggingface [OPTIONS] <COMMAND>

Subcommands: werk auth huggingface login, werk auth huggingface status, werk auth huggingface logout.

Also accepts the global options.

werk auth api-key

Generate API keys for the OpenAI-compatible server

werk auth api-key [OPTIONS] <COMMAND>

Subcommands: werk auth api-key generate.

Also accepts the global options.

werk temp list

List temporary entries in the model store

werk temp list [OPTIONS]

Also accepts the global options.

werk temp purge

Remove temporary files from the model store

werk temp purge [OPTIONS]

Option	Meaning / parser default
--dry-run	Show what would be removed without deleting anything

Also accepts the global options.

werk temp path

Print the temporary-files directory

werk temp path [OPTIONS]

Also accepts the global options.

werk cache list

List local caches and saved chat histories by type and size

werk cache list [OPTIONS]

Option	Meaning / parser default
--kind <KIND>	Show only this storage type [possible values: chat-kv, chat-history, omlx-worker, runtime-state]
--json	Print machine-readable JSON

Also accepts the global options.

werk cache purge

Remove a listed cache, or all inactive caches with –all

werk cache purge [OPTIONS] <CACHE|--all>

Arguments:

[CACHE]  Exact cache ID from werk cache list

Saved chat histories are excluded from –all unless –include-history or –kind chat-history is supplied. Active or protected entries are skipped. Model weights are never removed.

Option	Meaning / parser default
--all	Remove all inactive, unprotected caches
--kind <KIND>	Limit –all to this storage type [possible values: chat-kv, chat-history, omlx-worker, runtime-state]
--include-history	Also delete saved chat histories with –all
--dry-run	Show what would be removed without deleting anything
--json	Print machine-readable JSON

Also accepts the global options.

werk runtime info

Show Werk Protocol and active-backend information

werk runtime info [OPTIONS]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode

Also accepts the global options.

werk runtime capabilities

Show truthful runtime capability statuses

werk runtime capabilities [OPTIONS]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode

Also accepts the global options.

werk runtime memory

Show live host and accelerator memory telemetry

werk runtime memory [OPTIONS]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode

Also accepts the global options.

werk runtime states

List persistent and volatile inference states

werk runtime states [OPTIONS]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode
--model <MODEL>	Filter by installed model ID
--tier <TIER>	Filter by state tier [possible values: vram, ram, disk, external]
--limit <LIMIT>	Maximum entries
--cursor <CURSOR>	Opaque pagination cursor from a previous response

Also accepts the global options.

werk runtime state

Apply an explicit action to one inference state

werk runtime state [OPTIONS] <ID> <COMMAND>

Subcommands: werk runtime state pin, werk runtime state unpin, werk runtime state promote, werk runtime state demote, werk runtime state evict.

Arguments:

<ID>  Opaque runtime-state ID

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode

Also accepts the global options.

werk runtime prune

Prune explicitly selected inference states; dry-run by default

werk runtime prune [OPTIONS]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode
--id <IDS>	Exact state ID; repeat as needed
--model <MODEL>	Select states for one installed model
--tier <TIER>	Select states in one tier [possible values: vram, ram, disk, external]
--older-than-unix-ms <OLDER_THAN_UNIX_MS>	Select states last accessed before this Unix millisecond
--all	Select every visible state
--confirm-all	Required acknowledgement when –all is used
--execute	Actually delete; otherwise only preview

Also accepts the global options.

werk auth huggingface login

Store a Hugging Face access token for gated model pulls

werk auth huggingface login [OPTIONS]

Option	Meaning / parser default
--token <TOKEN>	Hugging Face access token; omit to enter it interactively

Also accepts the global options.

werk auth huggingface status

Show whether Werk can find a Hugging Face token

werk auth huggingface status [OPTIONS]

Also accepts the global options.

werk auth huggingface logout

Remove the Werk-stored Hugging Face token

werk auth huggingface logout [OPTIONS]

Also accepts the global options.

werk auth api-key generate

Create an API keys TOML file for werk serve

werk auth api-key generate [OPTIONS]

Option	Meaning / parser default
--path <PATH>	API keys file to create; defaults to ~/.config/werk1112/api-keys.toml
--name <NAME>	Human-readable name stored next to the generated key [default: default]
--force	Overwrite an existing API keys file

Also accepts the global options.

werk runtime state pin

Pin a state against policy eviction

werk runtime state <ID> pin [OPTIONS]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode
--execute	Actually apply the change

Also accepts the global options.

werk runtime state unpin

Remove a state’s policy pin

werk runtime state <ID> unpin [OPTIONS]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode
--execute	Actually apply the change

Also accepts the global options.

werk runtime state promote

Promote a state to RAM or VRAM

werk runtime state <ID> promote [OPTIONS] <TARGET>

Arguments:

<TARGET>  Target memory tier [possible values: vram, ram]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode
--execute	Actually apply the change
--allow-experimental	Allow an experimental backend adapter

Also accepts the global options.

werk runtime state demote

Demote a state to RAM or disk

werk runtime state <ID> demote [OPTIONS] <TARGET>

Arguments:

<TARGET>  Target lower tier [possible values: ram, disk]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode
--execute	Actually apply the change
--allow-experimental	Allow an experimental backend adapter

Also accepts the global options.

werk runtime state evict

Evict one explicitly named state

werk runtime state <ID> evict [OPTIONS]

Option	Meaning / parser default
--deployment <DEPLOYMENT>	Exact deployment instance for runtime control in profile mode
--execute	Actually delete the state

Also accepts the global options.

This site is open source. Improve this page.