Gradio seems to double VRAM requirement

[dzupin]

Running sora examples in CLI: python inference.py --save_dir=output/sora/ requires about 12.7 GB VRAM

Running same examples in gradio GUI: python app.py requires more than 24 GB VRAM (fails to fit into my 4090 card)

There seems to be something wrong with how memory is used in Gradio app