In the gradio_app/app.py:
1- load one model at a time,
2- c_dtype = torch.bfloat16
3- miss a generator=generator on def generate_prior(...)
as an example see: https://github.com/another-ai/stable_cascade_easy
With a nvidia rtx 3060 12 gb vram: 10 minutes vs 44 seconds
In the gradio_app/app.py: 1- load one model at a time, 2- c_dtype = torch.bfloat16 3- miss a generator=generator on def generate_prior(...) as an example see: https://github.com/another-ai/stable_cascade_easy With a nvidia rtx 3060 12 gb vram: 10 minutes vs 44 seconds