Open YukiSakuma opened 2 years ago
I'm afraid you will ever get a response. That boris guy or his team keeps silent about requests of changing that webapp's mechanism from the beginning. And as far as I know, that webapp is not OSS so you can't even push some PR on it.
Since currently running dallmega interfacing in a non-multiGPU environment is nothing but a waste / nothing but a unnecessary ram hog, so the first thing to do had better be asking them to open that webapp's (or its backend) code and accepts PR I guess.
I'm afraid you will ever get a response. That boris guy or his team keeps silent about requests of changing that webapp's mechanism from the beginning. And as far as I know, that webapp is not OSS so you can't even push some PR on it.
Since currently running dallmega interfacing in a non-multiGPU environment is nothing but a waste / nothing but a unnecessary ram hog, so the first thing to do had better be asking them to open that webapp's (or its backend) code and accepts PR I guess.
Well the creator created a new website for demo (https://www.craiyon.com/) but sadly still no manual option to input parameters...
Hey, we cannot input custom parameters in the demo as we wouldn't be able to batch prompts. The issue is that we compile putting those parameters as static as you can see in the colab.
My recommendation is to use the colab where I made every relevant parameter accessible and added comments. It will recompile each time you change those parameters so it may add delays. Potentially they wouldn't need to be static but then you could need to ensure not to use default values and you will get some errors during compilation.
Sounds like a lousy excuse, there's nothing "static" in py world and they won't be "compiled" to anywhere. Maybe that just mean that webapp has a bad architect. (as no outsider can see its codes, either fully or partial)
Anyway, as I managed to make this thing run on free tier colab TPU without OOM recently, that webapp is not something I really care about now.
(To those who going to pay for a ColabPro due to megafp16's OOM, that's not worth it really) It can give me a 2MB+ seq dump after a full 12hr run. (so 3200+ images gen, as it decoded)
Here's the notebook, make sure the runtime type is 'TPU', then run both 'prep' and 'lokappz' section. DALL_EBackground.zip
https://huggingface.co/spaces/dalle-mini/dalle-mini In the demo inference site, It'd be nice to have a feature that is to be able to input your own parameters such as
seed
,gen_top_k
,gen_top_p
, for example for the sameprompt
andseed
I want to be able to compare the results for differentgen_top_k
,gen_top_p
values.