we did the same thing lol

By the way, the 768 default trigger for tiling may not be adequate for Qwen: a 768x768 image needs 8157.94 MB (even 512x512 needs 3098.51 MB). Of course, we could consider someone playing with Qwen has to deal with memory issues already, since generation reaches 10978.29 MB for a Q4_K_S quant...

I'm hoping leejet optimizes the VAE memory usage

@LostRuins , this isn't enough for img2img: I'm getting an assertion failure for images large enough to trigger the tiling:

ggml/src/ggml-backend.cpp:300: GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds") failed

Edit: fixed: #1785