GitLab

Do inference in fp32 to make sure quality stays the exact same.

Try to keep models on the vram when possible.

Better lowvram mode for controlnets.

load_model_gpu() is now used with the text encoder models instead of just
the unet.

This is faster on big text encoder models than running it on the CPU.

Add a new argument --use-quad-cross-attention

Make the CLIP model work on the GPU.