# Eng Adapt — Chunked ONNX CDN (Kokoro + FuguMT)

> Browser TTS + translation models on Cloudflare Pages (fp32, Brotli chunks).

## App

- Main app (offline PWA): /index.html
- Service Worker: /sw.js
- Offline plan: see repo v2/docs/offline-support-plan.md

## Documentation

- Integration guide: /HOW_TO_USE.md
- Machine-readable manifest: /kokoro-cdn.json
- Chunked fetch library: /lib/chunked-model-fetch.js

## Kokoro TTS

- model_id: onnx-community/Kokoro-82M-v1.0-ONNX
- dtype: fp32
- manifest: /models/onnx-community/Kokoro-82M-v1.0-ONNX/onnx/model.onnx/manifest.json
- chunks: 15 (Brotli, 20 MiB each)

## FuguMT translation

- model_id: Kadonox/fugumt-en-ja-onnx
- dtype: fp32
- license: CC BY-SA 4.0
- models_root: /models
- config: /models/Kadonox/fugumt-en-ja-onnx/config.json
- encoder manifest: /models/Kadonox/fugumt-en-ja-onnx/onnx/encoder_model.onnx/manifest.json (4 chunks)
- decoder manifest: /models/Kadonox/fugumt-en-ja-onnx/onnx/decoder_model_merged.onnx/manifest.json (4 chunks)
- compressed total: ~147 MB (encoder ~67 MB + decoder ~79 MB)

## Important warnings

1. ONNX weights are NOT single files. Use manifest.json + part-*.br per logical .onnx path.
2. Concatenate chunks in index order, decompress ONCE (not per-chunk).
3. Use configureHubMirrorChunkedEnv() from /lib/chunked-model-fetch.js with modelsRootUrl = {origin}/models.
4. Do NOT fetch Hugging Face Hub URLs when same-origin models are available.

## Regenerate models

```bash
cd v2/cf-pages
node scripts/download-models.mjs [kokoro|fugumt|all]
node scripts/prepare-public.mjs [kokoro|fugumt|all]
```