African-Languages-Lab/all-lab-tts-endpoint

ALL Lab TTS — Inference Endpoint (custom Docker container, 38 languages) # ALL Lab TTS — Inference Endpoint (custom Docker container) Router package for a **Hugging Face Inference Endpoint**. Serves all 38 languages in the `Individual-TTS-Best-HumanEval` collection (`all-lab/all-lab-tts-{language}`) — see `handler.py`'s `MODELS` dict for the exact list. This is **not** a Space. Deploy with **Inference Endpoints → Custom Container**. ## Repo layout | File | Role | |------|------| | `Dockerfile` | Custom IE container (uvicorn on port **80**) | | `app.py` | FastAPI: `GET /health`, `GET /`, `POST /` | | `handler.py` | Multi-language OmniVoice load/generate router (`MODELS` dict = the 38 language repos) | | `alllab_tts/` | Vendored inference package | | `requirements.txt` | Pins `transformers==5.13.0` (needs `HiggsAudioV2TokenizerModel`) | ## Image build (automatic) `.github/workflows/build-and-push.yml` builds and pushes on every push to `main`. No local Docker needed. **Image:** `ghcr.io` **Package visibility:** must be **public** (Settings → Danger Zone → Change visibility) at `github.com` — otherwise the Inference Endpoint's image pull fails with a 401. ## Deploy (do this in the HF UI) 1. Go to huggingface.co (or **New Endpoint** from your HF account). 2. **Model repository:** any accessible repo works as the placeholder — this isn't actually used by a custom-image deployment (e.g. `all-lab/all-lab-tts-hausa`). 3. **Container type → Custom**: - **Image URL:** `ghcr.io` - **Port:** `80` - **Health route:** `/health` 4. **Hardware:** GPU, e.g. AWS `nvidia-t4` / `x1`. 5. **Scaling:** `min_replica=0` for scale-to-zero billing when idle. 6. **Secrets:** `HF_TOKEN` = a token with **read access to all 38** `all-lab/all-lab-tts-*` private model repos. 7. Optional env var: `ALL_LAB_DEFAULT_LANGUAGE=hausa` (which language warms at container start — defaults to `hausa`). …