Fix NameError in download_weights.py and add GPU memory cleanup in gradio_app.py

[fox3nox]

Summary

• Critical bug fix: download_weights.py referenced args.models inside main(), but args is only defined in the __main__ block — this caused a NameError whenever main() was called programmatically or imported. Fixed by passing models as an explicit parameter to main().
• Validation improvement: Replaced the bare assert for invalid model names with a proper ValueError and a descriptive error message.
• GPU resource management: gradio_app.py now handles torch.cuda.OutOfMemoryError separately with a user-friendly message and immediately frees VRAM via torch.cuda.empty_cache() + gc.collect(). A finally block ensures cleanup runs after every generation, preventing VRAM leaks between calls.
• Temp file cleanup: Partial/failed output files are now deleted on error instead of being left on disk.

Test plan

• Run python download_weights.py --models 720x720_5s — should download without error
• Import main from download_weights in a Python shell and call it directly — no NameError
• Pass an invalid model name (e.g. --models invalid) — should raise ValueError with a clear message
• Trigger an OOM error in gradio_app.py (e.g. very high resolution) — should log a clear message and not leave temp files behind
• Run multiple generations in sequence and verify VRAM is released between runs

🤖 Generated with Claude Code