search

borisdayma / dalle-mini

DALL·E Mini - Generate images from a text prompt
https://www.craiyon.com
Apache License 2.0
14.76k stars 1.22k forks source link

Is there a way for the mega model to work with kaggle? #204

Open YukiSakuma opened 2 years ago

YukiSakuma commented 2 years ago

I'm using the notebook provided but on kaggle and using the DALLE_MODEL = "dalle-mini/dalle-mini/mega-1-fp16:latest" but it will crash out of memory error Your notebook tried to allocate more memory than is available. It has restarted.

auwsom commented 2 years ago

I got it to work after upgrading to Pro for $10.. and then waiting a bit or restarting. https://github.com/borisdayma/dalle-mini/issues/207

YukiSakuma commented 2 years ago

I got it to work after upgrading to Pro for $10.. and then waiting a bit or restarting. #207

It could definitely work with Pro, but I'm searching for a way if possible for the mega model to work with kaggle with only 13 GB of RAM

YukiSakuma commented 2 years ago

It seems a higher RAM is really needed but I don't know what is the minimum, I was able to load the mega fp16 with CLIP in paperspace free with an 8 GB GPU but RAM capacity is 30 GB, image generation for a total of 9 images is about ~5 minutes. Now I'm curious what would be the minimum RAM and GPU VRAM for inferencing once there is an open source DALLE-2 to play with.

borisdayma commented 2 years ago

What if you don't load CLIP?

YukiSakuma commented 2 years ago

What if you don't load CLIP?

I tried but OOMs at this line:

from flax.jax_utils import replicate

params = replicate(params)

I am using this edited notebook: https://www.kaggle.com/code/apolinario/dall-e-mega I need the _do_init=False, without it it would OOM at loading the mega model because the official inference notebook provided has the _do_init as an unexpected keyword argument. I also added line below at the top import os; os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform" As provided here:https://github.com/saharmor/dalle-playground/issues/14