Fix: efficient log reading (last 5000 chars only)

Updated: scaled pipeline with real data (10K images, 5K LLM samples, checkpoint/resume support)

Update app.py: full training pipeline with real datasets

Add constrained decoding: force visual tokens after <video_start>

Rewrite: generation-only app, preload models, no auto-training

Update app.py: add video generation tab

Fix: remove every param from demo.load

Update app.py: auto-start training, file-based logging

Update app.py: better error handling, show_error=True

Upload app.py

initial commit