Training in Runpod environment

#22
by iiitaiii - opened

@TheBloke - thank you for creating the Runpod template. I was able to get it up and running, but couldn't figure out how to train it on my own data under the training tab. Since there is no Jupyter Notebook (like in the SD templates), there is no folder for me to upload the raw text files (or Json). Do you know how I might be able to do training in the Runpod environment?

Thanks in advance!

Ah yeah I haven't added Jupyter to it, because I didn't see a need given models can be downloaded in the UI. I didn't think about training data. It's a shame text-gen-ui doesn't provide an option of uploading a text file in the UI.

There's two fairly easy options for uploading files to the pod:

  1. Use SCP to upload. This is easy on Linux and macOS; a little harder on Windows. But you can Google for free Windows SCP clients like WinSCP.
  2. Use runpodctl to transfer a file from local to the pod, following this guide: https://github.com/runpod/runpodctl#how-to-transfer-data

Let me know if either of those work for you. If you get stuck let me know and I'll give some more explanation.

Thank you! I'll try that and will let you know if that works.

@TheBloke - I was able to upload the file, but in the UI under Training>Raw Text File field, it's failing to see the file. I think it's because the folder structure is not there (I had to recreate it), but I'm sure there are other dependencies that it needs to see...

To give some context, I just want to add a Lora training with additional text to create some 'expertise' on top on the LLM model you created.

I hope this makes sense - I'm pretty new to this...

Were you definitely in the right place? It should have been under /workspace/text-generation-webui/training/datasets

I just checked my template and the folders are there, so I wonder if you were in the wrong location?

image.png

ok - I was working off an older runpod template. I redeployed one with the latest (one-click) and it has the complete folder structure in it. thank you!

This comment has been hidden

Were you definitely in the right place? It should have been under /workspace/text-generation-webui/training/datasets

I just checked my template and the folders are there, so I wonder if you were in the wrong location?

image.png

Thank you for sharing, @TheBloke . I have gone through the chat logs regarding this issue, and it seems like I have encountered a similar problem. I have used runpodctl to send a file and received it in the terminal of TheBloke Local LLMs One-Click UI. However, I'm unsure how to transfer the file to "/workspace/text-generation-webui/training" so that it can be found in the dataset. Do I need to specify this path during the file reception process? Could you please guide me on how to proceed? Thank you!

I have successfuly upload the file to the path: ls /workspace/text-generation-webui/training/datasets
ALL.JSON put-trainer-datasets-here.txt
However, there is nothing showed in the webui training/dataset. I tried to refresh and restart, and it just does not work.
Is it because I used the wrong template?

@evan624 - I was able to get it working with the template (thebloke/cuda11.8.0-ubuntu22.04-oneclick:latest) -- however, I'm still getting an error when trying to run the training...

@evan624 - I was able to get it working with the template (thebloke/cuda11.8.0-ubuntu22.04-oneclick:latest) -- however, I'm still getting an error when trying to run the training...
Thanks!
I saw when someone uses the monkeypatch for GPTQ model loader, it will work.

Sign up or log in to comment