Merge the final LoRA finetuned checkpoint

[awaelchli]

Merges the final checkpoint saved by the lora.py finetuning script.

[rasbt]

Nice! In this case, we can now also simplify the docs right? I.e., at https://github.com/Lightning-AI/litgpt/blob/wip/tutorials/finetune_lora.md#merging-lora-weights

[awaelchli]

@rasbt Yes it's a nice improvement. But I need to be a bit careful here. Right now, merging a model requires it to fit into CPU RAM. For the 70B models, that could potentially be a problem on consumer hardware. I think it would be smarter to first implement sequential merging + sequential write (the latter we have already in LitGPT). That's why the PR is still in DRAFT.

[awaelchli]

I changed my mind. I think it would be good and safe to do this because the manual merging logic is the same. I also used this branch in my finetuning Studio.

[awaelchli]

Nice! In this case, we can now also simplify the docs right? I.e., at wip/tutorials/finetune_lora.md#merging-lora-weights

@rasbt Yes. It would simplify in the sense that we could immediately suggest the user to run generate or chat.

[carmocca]

It would simplify in the sense that we could immediately suggest the user to run generate or chat.

Do you want to do this in this PR?

[awaelchli]

@rasbt I simplified the merging section based on the changes in this PR. Would you like to review it and edit if necessary?

[rasbt]

This is nice! Just a few minor suggestions below.