model_name_or_path is not anymore in Generate() parameters (cli.py)

[MLator]

Hey @s9roll7

1. model_name_or_path
   The problem is it was very useful to load a specific model without having to load a secondary one in the config file, to fit in 12G Ram memory of colab (GPU and system).
   Many checkpoints have already all necessary data (TE, UNet, VAE, ...), so it was great and saving time at load also.

2. How much ram did you need for sdxl turbo example ? (system ram and GPU ram)
   As for me it crashes (full system memory) in sdxl_models/unet.py during call of
   model = cls.from_config(config, **unet_additional_kwargs)

I didn't try yet but maybe it could be possible by the way to create the temporal_unet model on the GPU ram (which is empty at this moment) and then after the 16bits conversion to move it back to system ram.

[Eye-01]

Hey,
I agree on point 1, I made such changes if you're interested as well as a few fixes to free mem earlier between system and GPU, I'll commit them soon on my fork and will propose a PR.