If I git pull will I have to remake the training weights?

[nazthelizard122]

I downloaded this project a while ago since it is the best interface with LLaMa at the moment, considering the weights are a few hundred GB total, it would be quite inconvenient to have to re-quantisize them every time I git pull, Is this an issue I should be concerned about?

[j-f1]

You can check the change history for quantize.cpp and convert-pth-to-ggml.py — unless those are changed, you probably don’t need to recreate the ggml files. The most recent change to those files is optional to re-apply but it improves output quality for non-Latin scripts.

[nazthelizard122]

Thanks!