VOCENCE_HF.md

# Publishing **Sad-tts-v1** (`KGSS/Sad-tts-v1`)

1. **Weights** — Full OmniVoice-style snapshot (main `model.safetensors` + `audio_tokenizer/`). Upstream reference: **[k2-fsa/OmniVoice](https://huggingface.co/k2-fsa/OmniVoice)**.

   - `model.safetensors`
   - `config.json`, `tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`, etc.
   - `audio_tokenizer/` (directory with its `model.safetensors` and configs)

   You can sync without Git LFS clutter locally:

   ```bash
   huggingface-cli download k2-fsa/OmniVoice --local-dir ./mysnap
   cp -r miner.py chute_config.yml vocence_config.yaml README.md ./mysnap/
   # then push ./mysnap to your Hub repo
   ```

2. **Revision** — Pin the commit SHA in your deploy script as `VOCENCE_REVISION` and use the same value in `vocence miner commit --model-revision`.

3. **`chute_config.yml` on SN78** — Subnet Chutes integrations require TEE + supported GPU class (see your working `rhy-TTS-v1` deploy); this folder’s `chute_config.yml` matches that pattern.

4. **Safetensors metadata** — OmniVoice checkpoints from the official Hub use standard PyTorch safetensors metadata; no `pt-sn78v`-style patching should be needed (unlike some custom Qwen repacks).