Skip to content

Image and video nodes ​

A machine with image or video models installed can run jobs for an orchestrator elsewhere. The orchestrator takes such nodes only when started with DPPAN_VISUAL_NODES=on; without it, dppan join --visual is refused and says so. Then, on the node:

bash
dppan join --orchestrator http://192.168.1.10:8001 --visual

Trusted networks only

Any machine that can reach the orchestrator's gateway port (9001 by default) can join as a node. A node is sent the prompts and input files of the jobs it runs, and what it returns is published as the result. Turn DPPAN_VISUAL_NODES on only where you trust every machine that can reach that port, or limit it with a firewall.

--visual offers every image and video model installed there; add --model to serve a text model as well. The orchestrator lists each model with the machines that hold it, in visual.places of GET /v1/generations/capabilities, and the nodes under GET /nodes/visual.

Where a job runs. A request's node names a place: local, or a node's name or id. A place that is not connected, or lacks the model, is refused with bad_params. Without node, the job takes the first free machine with the model, in this order: nodes serving only image and video models, the least busy first; then this machine; then nodes that also serve a text model. DPPAN_VISUAL_PREFER_NODES=off puts this machine first. Machines run at once, so two nodes make two videos together.

A node serving a text model as well (--model with --visual) keeps memory free for that model's chat: each image or video model there leaves room for two chat sessions on the layers the node serves, plus 1 GB, at most half the memory its model runs in (the GPU's). dppan join prints how much. A shorter DPPAN_CTX on the node makes each session smaller. To choose the amount yourself, or to run more chats at once there, set DPPAN_VISUAL_MAX_VRAM_GIB on the node.

Files. Input images and videos go to the node with the job, and the result comes back here, so a job looks the same wherever it ran. Its node_name says which node made it.

When a node leaves mid-job, a job that named no place moves once to another machine with the model. Until that machine is free, the job is in the waiting stage and its node_name is empty. One that named the node, or has nowhere to go, fails with node_left.

A node running an older dppan than the orchestrator may not read a newer job. A job that named no place then moves once to another machine, as above; one that named the node fails with node_outdated, naming the node and its version. Update dppan on the node.

A model the orchestrator also holds must have the same files on the node, or the node's offer is refused with the reason, which dppan join prints.

Free to run · Proprietary