76 lines
2.8 KiB
Markdown
76 lines
2.8 KiB
Markdown
# QVAC runtime and GPU authority
|
|
|
|
## The single-master rule
|
|
|
|
Only `daemon/qvac-master.js` may import and operate the QVAC SDK for Jarvis
|
|
runtime work. Skills receive a master facade; they do not call `loadModel()` or
|
|
create their own workers. The harness is connected through
|
|
`daemon/harness-bridge.js` and uses the same owner.
|
|
|
|
```mermaid
|
|
flowchart LR
|
|
V[Voice] --> M[Master scheduler]
|
|
H[Harness] --> M
|
|
CU[Computer vision] --> M
|
|
J[Jobs: media / RAG] --> M
|
|
M --> L{GPU admission}
|
|
L -->|fit| SDK["QVAC SDK 0.19.1"]
|
|
L -->|no fit| E[truthful unavailable error]
|
|
SDK --> W[one QVAC worker]
|
|
```
|
|
|
|
The optional `qvac serve --openai` command is diagnostic/API infrastructure.
|
|
It is not started alongside `jarvisd` in production and must remain bound to
|
|
`127.0.0.1` when used.
|
|
|
|
## Lifecycle
|
|
|
|
1. Read `QVAC_CONFIG_PATH` and the selected profile.
|
|
2. Query GPU visibility and model fit before loading the local chat model.
|
|
Opt-in Groq agent inference skips this local chat load; ASR and TTS still
|
|
use auxiliary QVAC models on the same worker.
|
|
3. Admit work through interactive, computer-vision, background, or
|
|
3. Admit work through interactive, computer-vision, background, or
|
|
maintenance lanes.
|
|
4. Load one model configuration at a time and serialize worker access.
|
|
5. Stream completion or media progress to the daemon.
|
|
6. Cancel by request ID/model ID when the user says stop or a grant is revoked.
|
|
7. Suspend or unload according to idle policy; close once during shutdown.
|
|
|
|
```mermaid
|
|
stateDiagram-v2
|
|
[*] --> Uninitialized
|
|
Uninitialized --> Preflight: start
|
|
Preflight --> Ready: GPU visible + fit accepted
|
|
Preflight --> Unavailable: GPU missing / fit rejected
|
|
Ready --> Loading: admitted request
|
|
Loading --> Loaded: loadModel
|
|
Loaded --> Busy: completion / media / vision
|
|
Busy --> Loaded: stream complete
|
|
Busy --> Loaded: cancellation
|
|
Loaded --> Suspended: idle policy
|
|
Suspended --> Loading: new request
|
|
Loaded --> Closing: shutdown
|
|
Suspended --> Closing: shutdown
|
|
Closing --> [*]
|
|
```
|
|
|
|
## Configuration
|
|
|
|
The repository config is `qvac.config.json`. Installed user sessions use
|
|
`~/.config/jarvis/qvac.config.json`, generated from
|
|
`packaging/qvac.config.template.json`. The server is localhost-only and the
|
|
default alias is `jarvis-master`. Model cache files belong under the user's
|
|
cache directory, never under a system-wide writable location.
|
|
|
|
## GPU policy
|
|
|
|
`JARVIS_GPU_REQUIRED=1` is set by the service. The master requests GPU device
|
|
and maximum GPU layer offload. An explicit GPU request (or
|
|
`JARVIS_GPU_REQUIRED=1`) fails directly on a GPU load error rather than loading
|
|
a CPU copy that the master would reject. Optional `device: 'auto'` callers outside
|
|
that policy can still fall back, and runtime status labels that backend as CPU.
|
|
Run
|
|
`npm run gpu-doctor` and consult [hardware compatibility](hardware-compatibility.md)
|
|
before changing a model profile.
|