Fix parameter naming issue with Mistral 7B v0.3

[davmacario]

See #1431 and #1444.

After some investigation, I noticed two things:

1. The current download script (litgpg/scripts/download.py), when using safetensors models, does not download the file model.safetensors.index.json (which contains the mapping between the parameter name and the corresponding chunk); for Mistral 7B v0.3, this file contains the "correct" parameter names, i.e., the ones that LitGPT is able to convert without issues.
2. The issue about parameter naming happens when attempting the conversion from the file consolidated.safetensors (which gets 'converted' to consolidated.bin in download.py); turns out this file contains the model parameters saved with a different name (I suspect it was created from the original Mistral AI implementation), while the files model-0000X-of-00003.safetensors are created from the HF Transformers "MistralForCausalLM" class, so they have the "correct" weight names.

By downloading and using the model.safetensors.index.json file in the same way as the pytorch_model.bin.index.json file when working with '.bin' models from HF, the consolidated.bin file is not used (as the mapping is only necessary for the files model-0000X-of-00003.safetensors) and the model can be downloaded and converted without issues.

I was also considering preventing the download of consolidated.safetensors for what I said above.
The huggingface_hub.snapshot_download function used in download.py accepts the arg ignore_patterns, which I would use to ignore this file.

[rasbt]

Thanks for the PR! When I understand correctly, the changes you are suggesting would only be in litgpt/scripts/convert_hf_checkpoint.py? Unfortunately, it looks like there were some accidents in your rebasing, so it looks like lots of duplicated changes got committed: There are currently more than 100 files changed. Would you mind rebasing again off of the recent main branch or maybe doing a fresh PR with just the files to be changed, e.g., litgpt/scripts/convert_hf_checkpoint.py? Let me know if you need help with that.

[davmacario]

Yep, sorry, I just realized this :) I was working on an older fork and made some mistake in syncing it to this specific branch (mistral-v0.3).
Will do ASAP!

[rasbt]

Thanks! Btw GitHub now has a useful "sync fork" feature that you can use in your forked repo

[davmacario]

Yep! I believe it was the culprit of the extra commits, though, as it default synced my fork's branch to the main branch without asking :/
I synced to this branch using remotes, and there were no issues.

I'm opening a new PR now!

[Andrei-Aksionov]

I prefer to sync my fork via GitHub actions CLI tool (gh) whenever I want to create a new branch.
Speeds up a process a little bit.