Bagheera
Bagheera
do you have training logs handy? wandb project would be nice to see
you might also want to try generating [a flame graph](https://www.brendangregg.com/flamegraphs.html) on a Linux system.
i think this problem is just exhibited very strongly in DeepFloyd samples and most likely impacts every single model on MPS (with varying levels of impact)
yes, in the deep-floyd/IF project we see these; https://github.com/deep-floyd/IF/blob/develop/deepfloyd_if/model/gaussian_diffusion.py#L739 but i can't remember anywhere seeing them in the diffusers project
the specifics of that i'll have to wrap my head around but the initial idea of decoupling the logic inside `__call__` so it can be more effectively monkeypatched downstream sounds...
there's the concept library on the hf hub from back in the day. for the uninitiated, it is/was a collection of dreambooths others had done, to make it easier to...
or i guess a scheduler wrapper that takes in its own callbacks, in teh case of SD.Next
that's a good point vlad. i was just thinking a preliminary attempt at a scheduler wrapper might result in some lessons being discovered that might help make a better upstream...
what a neat way to make use of the knowledge already in the model!
thanks for the cool feature @linoytsaban did you try anything with a few images or a large set to see if it scales at all? or does it require a...