# Omni comprehension setup

Open **Generate → Audio Analysis → Set up Omni**. The same setup is available
from **Video Analysis** and **Audio/video Analysis**; these analyze supplied media
files rather than capturing a live feed. The setup modal shares the generative
model module pattern: choose a model module, inspect the exact tag and download
size, select an offered accelerator, then confirm **Pull and deploy**. This is
an administrative operation. Merely opening the page performs inventory checks;
it does not install or load a model.

## What gets installed

The guided setup uses
[`robit-man/qwen-omni-adapters`](https://github.com/robit-man/qwen-omni-adapters/tree/54bca6ba5883285c4b6073bb217b0d2c800eda74)
at revision `54bca6ba5883285c4b6073bb217b0d2c800eda74`. It prepares the adapter's
Python environment, builds its pinned llama.cpp, pulls the selected trained Omni
audio bridge, and starts the core local comprehension and speech stack. It does
not invoke the upstream general-purpose deployment script or enable its public
tunnel, camera/microphone harnesses, pointing stack, or startup generation smoke.

| Model module | Exact Ollama tag | Approx. artifact size | Discrete GPU reservation |
| --- | --- | --- | --- |
| Ornith 1.5 9B | `robit/ornith-1.5-omni-audio-bridge:q4km` | 8.15 GiB | 24,000 MiB |
| Qwen3.8 27B E03 | `robit/qwen3.8-27b-e03-obliterated-omni-audio-bridge:q4km` | 18.33 GiB | 45,000 MiB |

Artifact size describes the download, not the complete runtime memory footprint.
Discrete NVIDIA deployments require Linux, a working user systemd bus, CUDA build
tools and the GPU broker. The selected full GPU UUID must pass the broker's
current health and admission checks. Holds, quarantined devices and transitional
leases remain authoritative. Supported Tegra hosts use the adapter's explicit
direct-device admission and kernel residency checks when no broker is installed.
An unavailable broker is not permission to bypass it. Setup never substitutes
CPU or another accelerator.

## Progress and cancellation

The daemon returns a durable job identity and a background worker reports
preparation, build, download, deployment, verification and rollback phases.
Progress stays in the setup modal. Closing and reopening the modal, or restarting
the daemon, reattaches to the existing job. The page retains a polling path if
event notifications are interrupted.

**Cancel setup** requests shutdown of this job's work. **Stopping** remains visible
until its process groups, owned user service and GPU lease have settled. Cached source,
builds and downloaded weights may remain. Failure to prove shutdown is an error,
not a successful cancellation. A cancellation arriving after the final verified
configuration commit reports that completion is finishing instead of promising
to undo it.

Readiness requires separate adapter-contract, owned-runtime and actual
accelerator proofs. A healthy HTTP response alone does not prove model residency.
Shared analysis configuration is saved only after the new runtime passes those
checks; a concurrent operator configuration change is preserved.

## Attach an existing stack

A healthy existing adapter offers **Use existing adapter** instead of replacement.
If no explicit model is saved, choose the matching bridge model module, then
confirm the exact tag shown. Adapter health does not identify loaded weights.
Attachment validates the adapter contract and saves the selected configuration;
the receipt keeps runtime and accelerator verification separate. Attachment does
not adopt, stop or restart that service.

Explicit `OMNIUS_OMNI_MODEL`, inherited `OMNI_OLLAMA_MODEL` and
`OMNIUS_OMNI_ADAPTER_URL` settings remain authoritative. Conflicting overrides
must be resolved before saving a different choice. Stored credentials remain
private and scoped to their saved endpoint/model. An explicit
`OMNIUS_OMNI_ADAPTER_TOKEN` can authenticate the same endpoint while its model is
selected deliberately.

## Storage and diagnosis

All paths are on the daemon's host. The normal configuration root is the user's
home; `OMNIUS_CONFIG_HOME` relocates it.

| Data | Location / ownership |
| --- | --- |
| Shared analysis selection | `.omnius/omni-analysis.json`, owner-only; read afresh by TUI and daemon comprehension requests |
| Runtime checkout and builds | `.omnius/runtimes/omni/source-<revision>/` |
| Derived speech cache | `<checkout>/runtime-data/components/tts-model.gguf` and `tts-projector.gguf`; tensor-bearing copies materialized from the selected bundle |
| Jobs, ownership and receipts | `.omnius/runtimes/omni/jobs/` and runtime metadata |
| Private bounded command diagnostics | `.omnius/runtimes/omni/setup-<job UUID>.log`, at most 4 MiB |
| Owned user service | `omnius-omni-<runtime-root hash>.service` |
| Downloaded weights | The local broker's shared Ollama model store |
| Generated media | The project that invoked generation; setup does not relocate it |

The derived speech files consume additional disk space beyond the original
combined-tag weights in the shared Ollama store. Setup retains these copies
after cancellation or service shutdown for later retries; `OMNI_KEEP_CACHE=1`
does not make them links to the original tensors.

The plan shows actual configuration/cache paths and broker coordination details.
Public status contains bounded phase diagnostics and receipts rather than raw
subprocess output, credentials or access tokens. Command diagnostics remain in
the private operation log. This setup dialog does not delete downloaded Omni
weights from the shared Ollama store.

If the worker itself crashes, the job is marked interrupted and retained service
state requires inspection before retrying. A leftover admission, runtime or
configuration recovery guard fails closed with a diagnostic. Verify the recorded
owner has exited and inspect the owned unit, cgroup and broker lease before
manual recovery; never delete a living owner's guard. Ordinary daemon restarts
leave the separately supervised worker running.

Source verification uses isolated command, service and GPU fixtures plus real
owned worker/browser processes. These checks do not claim that a model was
downloaded, deployed or run on the operator's hardware.
