Giters

transformerlab / transformerlab-app

Experiment with Large Language Models

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Mac OS using mlx - what is the required "Adaptor Name" field in Plugin Config tab of Training Template?

prashantnayak-edu opened this issue · comments

commented

Thanks for your question. That training plugin uses MLX Examples which is training using LoRA. LoRA creates an adaptor file that works alongside the original model, to power the new end model. The adaptor name is the name of the LoRA file that should be created.

In the end of the script, (https://github.com/transformerlab/transformerlab-api/blob/main/transformerlab/plugins/mlx_lora_trainer/main.py) the adaptor is fused with the base model, so you don't end up interacting with the adaptor file in the end, but it is generated and we may give the the option to interact with it later, hence why we ask for the name.

When you use other trainers in Transformer Lab, such as the llama_trainer powered by Hugging Face's Trainer library, the adaptor is kept separate entirely (not automatically fused) so it's more important for the other trainers that a helpful name is chosen.