Steward Garcia

@Green-Sky If it seemed easier to me to implement it this way (stream endpoint like chat), since otherwise it would have required websockets or a loop calling an 'http:127.0.0.0:7680/progress' endpoint,...

@Green-Sky ~~Unfortunately, I cannot run tests on CUDA Toolkit 11.8; I have no means to conduct the tests. I tried using Google Colab, but they already use the latest version...

@Green-Sky try now

Try enable SD_CONV2D_MEMORY_EFFICIENT this reduces the vae memory usage, or enable VAE tiling manually on the ui

@Green-Sky Thank you for fixing the error, and yes, I have been inactive because I've been feeling a bit demotivated.

Good advice!, We should take care it

It seems to be an error in the way matrix multiplications are performed in ggml. Does it work if you do it only with CPU?

@XienXX cmake .. - DSD_CUBLAS=OFF

I understand, for now the project is a mishmash of things, it still doesn't have a specific format to follow, the latest refactorings have made the code very verbose and...

I have a branch with the latest changes from ggml #221