finished the vae workers, forgot to add tiling support, all done. there won't be a resolution limit
tiling kicks in if height or width are greater than 1024, if i had more gpus i could increase the tiling threshold but i figure most people use tiling after 1024 anyway
for 1024 i only need 1gb for encode and 1.3gb for decode which is great, pleased with that, decode was at like 1.6gb, more fusions reduced it. iirc pytorch needs like 5gb for decode
next thing is for clip text i need to add a task to get pooler output or text embeds instead of last hidden state, easy enough
after that i think i can start getting everything deployed