What is this new "cat" architecture?

#1
by Winnougan - opened

I'm excited and curious about what improvements or differences "cat" architecture has over "pig" architecture. Can you elaborate?

What models we can use this?

cat encoder takes the tokenizer straight from the safetensors instead of rebuilding it; it solves some incompatibility issues between the text generation model and encoder purpose clip

for gemma2-2b, use it for lumina; for umt5xxl, use it for all wan video models, i.e., wan-s2v, wan2.2, wan2.1, etc.

Sign up or log in to comment