#Kohya_ss
21 messages · Page 1 of 1 (latest)
Kohya doesn't do inference, it's only for training
Non serverless, it has a gradio GUI
Do you know why there's no option to Save a configuration? I don't know if this is a necessary limitation, or a bug, or?
It gets disabled when you run it with --headless, which is required for it to work on RunPod without running into TKinter $DISPLAY errors, I will ping the developer and ask him if there is a possible work-around for it.
Thank you!
The next major release of Kohya_ss will have support for SDXL and will also fix the issue with the button to SAVE a configuration being missing when running with --headless.
"git checkout sdxl to test it out. Be aware that npz files created under sdxl will need to be deleted as they are not backward compatible with the previous release."
I tested the sdxl branch and can confirm that the "Save" button is now working on that branch. Now we just need to wait for it to be merged into the master branch, but feel free to checkout the sdxl branch in the mean time. Its awesome that he is working on sdxl support as well.
Jupyter isn't starting. I'm getting this error. I'm a noob. Should I delete the pod and try redeploying?
nohup: failed to run command 'jupyter': No such file or directory
[1]+ Exit 127 mkdir -p /workspace && cd / && nohup jupyter lab --allow-root --no-browser --port=8888 --ip=* --ServerApp.terminado_settings='{"shell_command":["/bin/bash"]}' --ServerApp.token=${JUPYTER_PASSWORD} --ServerApp.allow_origin=* --ServerApp.preferred_dir=/workspace &> /workspace/logs/jupyter.log```
Eh, I pip installed notebook and jupyterlab and then I was able to start jupyter. It's working for me now.
Woah... this version of Kohya_SS looks awesome.. settings presets? SDXL support? Yeah!
Thank you for letting me know, I'll look into it.
Jupyter has been fixed in the new 1.0.6 image.
Would be amazing to see a serverless/api implementation. Any idea what sort of lift would be required for that?
Why do you want to use it with Serverless? It takes ax while to train and will be expensive to run in Serverless. It's better suited to GPU cloud.
Doesn't GPU cloud necessarily mean that I have to interact with it manually? I'm trying to make an API that can run automatically. But I don't need it remotely frequently enough to just keep a pod running 24/7. Long story short, I'm trying to come up with an on-demand automatic kohya LoRA training API in the cloud.
Kohya doesn't have an API though, but you could probably press the button to output the command and then call that command manually from serverless
Well, yeah, that's why I was asking about an API implementation 😅 I mean, Kohya doesn't have an API or a UI, it's a separate package that adds a UI.
This template has been updated to Kohya_ss v21.8.6.