- Migrate backend from REST API to Gradio @gradio/client for generation
- Fix Gradio parameter alignment (positions 36-49) for reference/cover audio
- Add LoRA training pipeline with dataset upload, preprocessing, and export
- Add News page with dismiss/restore and GitHub star button
- Add localization info icon in Settings language section
- Fix upload audio URL prefix, add missing MIME types
- Add training API routes and Python preprocess script
- Update i18n with news keys for all languages
Phase 4.3: Largest component change. Adds cascading genre picker using
Phase 3 style data, DiT model selector with backend sync, EditableSlider
for all numeric parameters, LoRA load/unload panel, bulk generate, vocal
gender selector. Batch size max kept at 4 (not 8), inference steps max
kept at 32 (not 200) to prevent OOM. Added LoRA API stubs to api.ts.
Added ditModel to GenerationParams type.
Phase 4: i18n all 15 components from PR #19. Sidebar gains collapse/expand
toggle. Player gets playback speed selector (0.25x-2.0x) fixing bug B1
(hardcoded Chinese '正常'). SongList adds inline title editing and model
version badge. Delete actions now use ConfirmDialog instead of
window.confirm. Volume persisted to localStorage. Training nav item
deferred to Phase 7.
- Add lmModel parameter through full chain (types, API, routes, service)
- Add LM Model dropdown in Advanced Settings (defaults to 0.6B)
- Pass lmModel and lmBackend to format/enhance endpoint
- Update format_sample.py to accept --lm-model and --lm-backend args
- Auto-download model from HuggingFace if not present locally
- Persist model selection in localStorage
- Improve format route error logging with exit code and stdout/stderr
Fixes#9
fspecii