CLI reference
Inclavate installs three programs:
| Program | What it does |
|---|---|
dppan | The command you use: download models, join a node, run a model locally, start the dashboard |
dppan-orchestrator | The server that takes requests and sends work to nodes. Its flags are all settings |
dppan-node | The node server. dppan join starts and registers one for you; you rarely run it yourself |
Every program also takes --help, --version and --no-banner (or INCLAVATE_NO_BANNER=1) to hide the startup banner.
dppan
dppan [OPTIONS] <COMMAND>| Option | Default | |
|---|---|---|
--log-level | warn | trace, debug, info, warn or error. Also LOG_LEVEL |
--config <FILE> | ~/.dppan/config.toml | Where dppan keeps its settings, such as the orchestrator dppan join last used |
| Command | |
|---|---|
join | Join the network as a node |
pull | Download a model, or part of one |
run | Run a model on this machine alone, with no orchestrator |
ui | Start the dashboard |
status | Show the orchestrator and its nodes |
models | List the model catalogs and what is installed |
download | Download a model through Ollama |
manifest | Make a per-tensor integrity manifest from a full model file |
bench | Measure this machine's speed |
update | Update to the latest release |
uninstall | Remove the install and its PATH entry |
dppan join
dppan join --orchestrator <URL> [OPTIONS]| Flag | Env | Default | |
|---|---|---|---|
--orchestrator | ORCHESTRATOR_URL | required | Orchestrator URL, e.g. http://192.168.1.10:8001 |
--model, -m | DPPAN_MODEL | llama3.2:1b | Ollama model name or path to a .gguf file. With --visual and no --model, the node serves no text model |
--gguf | GGUF_PATH | none | A model file already downloaded (a full file or a dppan pull shard); skips Ollama. --model stays the id it registers with |
--layers | auto | Layer range to serve, inclusive both ends like pull (e.g. 0-15 = 16 layers). A shard defaults to its own range | |
--visual | off | Also offer this machine's image and video models: see Image and video nodes | |
--name | from device, model and layers | A name for this node, shown in the dashboard | |
--device | auto | auto, metal, cuda or cpu | |
--gateway-url | GATEWAY_URL | the orchestrator's host, port + 1000 | Where to dial the orchestrator's gateway, e.g. http://orch:9001 |
--port | PORT | 5001 | The node's HTTP port |
--grpc-port | GRPC_PORT | 5002 | The node's gRPC port |
--legacy | off | Listen on inbound ports instead of dialling the orchestrator; those ports must then be reachable | |
--http-url, --grpc-url | http://localhost:<port> | The addresses other machines use to reach this node's ports |
A node joins through the orchestrator's gateway and needs no inbound ports. dppan join remembers the orchestrator in ~/.dppan/config.toml, so dppan status and dppan ui find it without --orchestrator.
dppan pull
dppan pull <MODEL> [OPTIONS]| Flag | Default | |
|---|---|---|
<MODEL> | required | llama3.3:70b (Ollama) · hf:owner/repo[:QUANT] · inclavate:slug |
--layers | all | Layer range to download, inclusive (e.g. 0-26): see Sharded download |
--orch | off | Include the orchestrator's tensors, and download a published projector |
--projector-only | off | Download only the orchestrator's projector |
--projector <FILE> | auto | Pick an exact projector instead of the automatic F16, then BF16 |
--projector-out <PATH> | model store | Where to write the projector |
--allow-unverified-projector | off | Accept a projector whose source publishes no SHA-256, skipping that check |
--out, -o | ~/.dppan/models/… | Where to write the shard, or the model's directory for speech and music |
--verify | off | Check every tensor's SHA-256 against the integrity manifest |
--manifest | auto | Manifest file or URL for --verify |
--remote | off | Fetch byte ranges even when the full file is already here |
--speech | off | Pull a speech model: a backbone and its audio decoder. For a repository that publishes a PyTorch checkpoint, this converts it: see Speech › Getting a model |
--speech-decoder <MODEL> | Where the speech decoder lives, when it is not published beside the backbone | |
--f5-variant <v0|v1> | Which generation an F5-TTS model is, when nothing else says | |
--component <ROLE> | Pull one part of a model published as several files. Music models are assembled this way: see Music › Getting a music model | |
--f16 | off | Store an unquantised component's F32 tensors as F16: half the size, roughly half the generation time. Lossy, so opt-in |
--visual | off | Install an image or video model from the catalog by its id, e.g. dppan pull --visual z-image-turbo: see Images › Getting an image or video model |
--from <DIR> | With --visual: look in this folder first, and link in any file whose size and SHA-256 match instead of downloading it. Repeatable |
For speech and music models, pass --out and point every component of one model at the same directory: that directory is the model. Projectors are kept only on the orchestrator, under ~/.dppan/models/projectors/<model-id>/; a projector written elsewhere with --projector-out needs projector_path in that model's config/models.toml entry.
dppan run
dppan run [OPTIONS]Runs a whole model on this machine, with no orchestrator: interactive, or one prompt with --prompt.
| Flag | Default | |
|---|---|---|
--model, -m | llama3.2:1b | Ollama model name or path to a .gguf file. Also DPPAN_MODEL |
--prompt, -p | Answer one prompt, then exit | |
--max-tokens | 512 | Most tokens to generate |
--temperature | 0.7 | Sampling temperature; 0 always picks the likeliest token |
--top-p | 0.9 | Nucleus sampling threshold |
--device | auto | auto, metal, cuda or cpu |
--think, --no-think | the model's own | Turn a thinking model's reasoning on or off |
--reasoning-effort <LEVEL> | How hard to think: none, low, medium, high, as the model allows | |
--image <PATH> | Attach an image. Repeatable | |
--audio <PATH> | Attach a WAV or MP3 file. Repeatable | |
--video-frames <PATH@MS> | Attach one video frame and its time in milliseconds. Repeatable, in order | |
--projector <FILE> | the one pull downloaded | Use this projector instead |
More: Chat › From the command line.
dppan ui
| Flag | Default | |
|---|---|---|
--orchestrator | ORCHESTRATOR_URL, then the one dppan join remembered | Orchestrator URL |
--port | 3000 | Dashboard port |
--host | 127.0.0.1 | Bind address: 0.0.0.0 opens it to your network |
--open | off | Open the browser once it is up |
dppan status
| Flag | Default | |
|---|---|---|
--orchestrator | ORCHESTRATOR_URL, then the one dppan join remembered | Orchestrator URL |
--watch, -w | 0 | Refresh every N seconds; 0 prints once |
dppan models
Lists the curated catalog, and which models are already here.
| Flag | |
|---|---|
--tts | List speech models instead of text models |
--music | List music models |
--visual | List image and video models |
--json | Print the catalog as JSON |
dppan download
dppan download <MODEL> [--force]Downloads a model through Ollama, e.g. llama3.2:1b. --force downloads it again even if it is already here.
dppan manifest
dppan manifest <GGUF> [--out <FILE>]Makes a per-tensor integrity manifest from a full model file, or a local Ollama model id, for --verify. Prints it, or writes it to --out.
dppan bench
| Flag | Default | |
|---|---|---|
--model, -m | llama3.2:1b | Model to measure with |
--steps | 32 | Decode steps to run |
dppan update
| Flag | |
|---|---|
--version <VERSION> | Install this version instead of the latest, e.g. v0.8.5 |
--force | Reinstall even if already on the latest |
dppan uninstall
--yes (-y) skips the confirmation prompt.
dppan-orchestrator
dppan-orchestrator --port 8001 [OPTIONS]Every flag is also an environment variable, and every one is listed in Configuration. The ones most setups use:
| Flag | Env | Default | |
|---|---|---|---|
--port | PORT | required | REST API port |
--admin-token | DPPAN_ADMIN_TOKEN | none | Turns on admin endpoints and the dashboard's Admin tab |
--models-config | DPPAN_MODELS_CONFIG | config/models.toml (the current directory, else next to the binary) | The model registry |
--log-level | LOG_LEVEL | info | trace, debug, info, warn or error |
dppan-node
dppan join starts and registers a node for you. Run dppan-node yourself only to serve a fixed range of a model file:
| Flag | Env | Default | |
|---|---|---|---|
--gguf-path | GGUF_PATH | required | The model file |
--start-layer | START_LAYER | 0 | First layer served, inclusive |
--end-layer | END_LAYER | required | One past the last layer served |
--orchestrator-url | ORCHESTRATOR_URL | none (runs alone) | Orchestrator to register with |
--model-id | MODEL_ID | none | The model id to register with, e.g. llama3.2:1b |
--device | DEVICE | auto | auto, metal, cuda or cpu |
--port | PORT | 5001 | HTTP port |
--grpc-port | GRPC_PORT | 5002 | gRPC port |
--node-id | NODE_ID | generated | Node id |
--http-url, --grpc-url | HTTP_URL, GRPC_URL | http://localhost:<port> | The addresses other machines use to reach this node |