Troubleshooting and FAQ
GPU
- "CUDA error: the provided PTX was compiled with an unsupported toolchain": the NVIDIA driver is too old; update to 576.02 or newer.
cudart64_*.dll/libcudart.so.12not found: the CUDA builds carry these libraries next todppan, so the installation is incomplete. Run the installer again:curl -fsSL https://dl.inclavate.io/install.sh | sh -s -- --cudaon Linux,irm https://dl.inclavate.io/install.ps1 | iexon Windows. The CUDA Toolkit is not needed.- Falls back to CPU instead of GPU: run
nvidia-smi; the GPU must be Turing or newer (GTX 10xx and older are unsupported). - macOS GPU unused: use the
macos-arm64build on Apple Silicon (Intel Macs run CPU-only).
Chat
- A reasoning model's reply comes back empty. Its thinking is spent from the same
max_tokensbudget as the answer. RaiseMAX_TOKENSon the orchestrator (default 512), or the request'smax_tokens: see Reasoning. - A long conversation forgets how it began. Past
MAX_CONTEXT_TOKENS(default 2048), the oldest turns are dropped. Raise it, andDPPAN_CTXon the nodes with it: see Hardware and memory. - A whole video file is refused. Sending a video file needs
ffmpegandffprobeon the orchestrator'sPATH, or send frames instead: see Optional: ffmpeg.
Speech, music, images and video
- A transcript comes back empty.
languageis a filter, not a hint: a clip the model hears as another language comes back empty. Leave it out to keep whatever it detects: see Speech to text. - Speech runs on and on. See When a model will not stop.
/v1/audio/speech,/v1/audio/musicor/v1/images/…returns a job ID instead of the file. The job took longer than the endpoint waits: poll the job, or raise the wait under Configuration.429from a generation request: the queue is full; retry later.422: nothing in this deployment makes that modality with that model; retrying will not help.- A model you installed is not served. The orchestrator serves the models installed when it starts: restart it. Its startup log names each model it serves, and why it skipped any folder.
- An install asks you to update dppan. The catalog is newer than your dppan: run
dppan update.
Nodes
- A node cannot connect. It needs outbound TCP to the orchestrator's port (8001 by default) and its gateway port (9001, the port + 1000). Nothing has to be opened on the node.
dppan join --visualis refused. The orchestrator takes image and video nodes only when started withDPPAN_VISUAL_NODES=on: see Image and video nodes.- A job fails with
node_outdated. The node runs an older dppan than the orchestrator: update dppan on the node.
The dashboard
- Other machines cannot open it. It answers only on its own machine unless started with
dppan ui --host 0.0.0.0. - No Admin tab. Start the orchestrator with
--admin-token. - The banner fills a log file. It shows only in an interactive terminal; turn it off anywhere with
--no-bannerorINCLAVATE_NO_BANNER=1.
FAQ
Is anything sent outside my network? No. Every model runs on hardware you control. What each machine receives is under What stays on which machine.
Do I need a GPU? No. Everything runs on the CPU too, more slowly; a GPU is used when there is one. See Hardware and memory.
Do nodes need open ports or a public address? No. Nodes dial out to the orchestrator.
Can I use a model I already have in Ollama? Yes, as it is: dppan join with --model alone takes it from Ollama.
Can I reuse model files from another app, such as ComfyUI? For image and video models, yes: dppan pull --visual <id> --from <folder> links in any file that matches. See Getting an image or video model.
Where are models and results kept? Models in ~/.dppan/models (DPPAN_MODELS_DIR), results and uploads in ~/.dppan/artifacts (DPPAN_ARTIFACT_ROOT).
How do I update? dppan update; see Installation › Updating.
Which models can I run? See Models.