Replies: 1 comment 3 replies
|
it's not training the VAE - the pre-cache ensures it is unloaded before training runs. but ROCm is pretty tricky to get right, and isn't something i can really debug for users. I removed my AMD GPU - bugs will go uncaught more and more over time. |
3 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
So I finally figured out how to set up a working version of simple tuner on Linux for Radeon's but compared to other trainers is very slow about two times slower And way more susceptible to out of memory errors. I suspect it's something to do with vae training, which is something I never do in other software, but I can't find a way to turn it off here. Is there a way to turn it off and not train the vae?
All reactions