Create SoVITS checkpoint dirs before saving

Run SoVITS CPU training in a single process

Widen SoVITS cleanup to all resume dirs

Add SoVITS-only cleanup for fresh retrains

Add task stop controls for live training

Fix dashboard sync and CPU SoVITS workers

Show live training status and internal GPT logs

Always refresh training environment before runs

Force-patch GPT data module for worker 0

Avoid GPT shm workers on batch 4

Use file-system torch sharing for training

Add safe web file management

Persist and poll live training logs

Increase batch size slider limits

Adjust training labels and epoch limits

Use persistent bucket storage when available

Patch v2Pro speaker preprocessing for torchaudio

Add v2Pro speaker embedding preprocessing

Serialize preprocess for concurrent training

Fix v2ProPlus preprocess output dirs

Add version switching and agent API

Reset stale GPT checkpoints before training

Use one GPT dataloader worker

Improve trainer UI and add GPT dependency

Create GPT-SoVITS checkpoint directories

Fix GPT-SoVITS module search path

Fix Daniya GPT-SoVITS training space

fix: force rm -rf before clone, capture stderr

fix: use HOME/GPT-SoVITS instead of /tmp to avoid root-permission issues

trigger rebuild

fix: auto-clean corrupt GPT-SoVITS dir before re-clone

refactor v2: drop Dockerfile, lazy-install GPT-SoVITS at runtime, pin huggingface_hub<0.27

refactor: full rewrite with Dockerfile + Gradio 5.x

fix: add server_name=0.0.0.0 for HF Spaces compat

Add app.py