GitLab

Fix issue with not loading the SSD1B clip correctly.

support masked attention.

Use a simple CLIP model implementation instead of the one from
transformers.

This will allow some interesting things that would too hackish to implement
using the transformers implementation.