I’m currently fine-tuning LLAMA2 7b and came across discussions regarding whether to merge the base model with LORA weights. Some sources suggest merging, while others don’t.
Is merging necessary? What are the differences between the two approaches? And which one is preferable in terms of accuracy, inference time, and resource consumption?
From my understanding, merging seems essential because it combines the knowledge from the base model with the newly added weights from LORA fine-tuning. The base model holds valuable information, and merging ensures the incorporation of this knowledge with the enhancements introduced through LORA.
I’d appreciate insights from the community on whether merging is a recommended practice and its impact on model performance, inference speed, and resource utilisation.
Resource 1 (No merging): How to Fine-tune Llama 2 with LoRA for Question Answering: A Guide for Practitioners
Resource 2 (Merging) : How to Fine-tune Llama 2 With LoRA - by Derrick Mwiti