Release v2.10.1 · tenstorrent/tt-studio

🚀 New Features

External Model Registration: Register model containers that aren't in the catalog, and make eligible ones available to the Apps marketplace

Fixed valid deployment records being falsely marked as stopped, along with a batch of Voice Agent and deploy-progress issues found in prod mode:
- Voice Agent models now deploy sequentially instead of colliding on a deploy_in_flight 409, and the page shows download progress
- Prod mode runs a single uvicorn worker — four workers were splitting in-memory deploy state, so progress polls hit the wrong process and healthy deploys reported not_found; the deployment store also gained a per-process tmp file and fsync so concurrent writes can't corrupt deployments.json
- Benign HF_TOKEN log lines (including ✅ HF_TOKEN is valid.) are no longer classified as authentication failures
- Weightless models (Whisper, SpeechT5) use the full progress bar during image pull instead of sitting at 25%
- The Voice Agent no longer records forever when VAD misses speech onset, and .mjs is served with a JavaScript MIME type so the VAD runtime loads in prod
- Added check_hf_repos for per-repository Hugging Face access checks, so gated models are handled before deployment
Fixed deploy secrets being lost when deploys run concurrently
Fixed Docker Control supervisors accumulating and fighting over port 8002
Fixed the wakeword WebSocket not being proxied in nginx prod, which broke voice activation in production
Ungated models now deploy without a Hugging Face token saved
The footer shows all device IDs for multi-chip models instead of just the first

Refreshed CODEOWNERS for the current repo layout and review lanes
Granted write permissions to the community labeling workflow
Added qbge-devex-01 and qbge-devex-02 as selectable runner labels in the deploy healthcheck workflow

Full changelog → #1287