3 ms·
Yeah, high res performance is very non linear, especially without swapping out the attention for xformers, flashattention2 or torch SDP (and I don't think torch
by brucethemoose2 3y ago
Yeah, high res performance is very non linear, especially without swapping out the attention for xformers, flashattention2 or torch SDP (and I don't think torch MPS works with any of those).
That model doesn't work well at 1024x1024 anyway without some augmentations. You want this instead: https://huggingface.co/segmind/SSD-1B https://huggingface.co/segmind/SSD-1B