error: cannot unpack non-iterable ActionTypeHint object

[pvaldivia4]

Bug description

Dear Lightning,

I'm attempting to follow a medium tutorial using my own dataset. I have downgraded lightning due to a suggestion on issue #1915. On step 3 of the tutorial, the fine-tuning step using litgpt, I receive the message "error: cannot unpack non-iterable ActionTypeHint object".

I am on Version 0.5.7 of litgpt and have train.json as well as the checkpoints/mistralai/Mistral-7B-Instruct-v0.3 folders in my working directory

Thus far, I have tried:

• Reading the finetune_lora readme
• installing litgpt[all] as suggested in your documentation, since the medium article only installed litgpt
• modifying the command in the medium article according to the documentation i.e. using: litgpt finetune_lora checkpoints/mistralai/Mistral-7B-Instruct-v0.3 --data JSON --data.json_path train.json --data.val_split_fraction 0.2 --out_dir out/custom-model
• Creating the output directories [either finetuned (medium article), or out/custom-model (documentation)] ahead of running the commands

Appreciate any help that you can provide.

What operating system are you using?

Linux

LitGPT Version

Version: 0.5.7

[syntheticgio]

Fwiw, I'm seeing the same thing on a fresh install of litgpt. Also trying to follow the tutorial. Same version. I'm going to see if I can downgrade and try again

Going back to 3.1 I'm still seeing it (for finetune, not finetune_lora since that wasn't an option there).

litgpt finetune_lora stabilityai/stablelm-base-alpha-3b --data Alpaca
Does work, something to do with the data. I've tried the same as the OP with the data file being what was shown in the tutorial here: https://github.com/Lightning-AI/litgpt/blob/main/tutorials/finetune_lora.md#tune-on-your-dataset

data file:
[
{
"instruction": "Arrange the given numbers in ascending order.",
"input": "2, 4, 0, 8, 3",
"output": "0, 2, 3, 4, 8"
}
]

[syntheticgio]

This appears to be something related to jsonargparse library - as far as I can tell nothing is wrong in the code although I'm not familiar with this particular parser; I'm going to try a different version of python; I'm using 3.12 currently.

[syntheticgio]

@pvaldivia4 changing to using python3.10 worked for me, I think jsonargparse needs to be updated to work with 3.12 (maybe also 3.11, not sure which one broke it).

Assuming you have python3.10 installed:
python3.10 -m venv venv
source venv/bin/activate
pip install -e '.[all]' (you could probably install the regular pip install litgpt[all] but this is what I did.

I also created a directory for the data files, and created two files in there - train.json and val.json; for the --data.json_path I pointed to that directory. You can do it with a file, I think, but you might have to specify val_split_fraction if you do it that way.

If I see a version of jsonargparse that is compatible with python3.12, I'll open a PR.

[syntheticgio]

@pvaldivia4 if you can't downgrade or install a different python version, I've made a branch that should work https://github.com/syntheticgio/litgpt/tree/add-compatibility-for-python3.10%2B

python -m venv venv
source venv/bin/activate
pip install -e '.[all]'

I'll make a PR request, but not sure what the maintainer's process is for accepting those or if they are willing to try to support python3.12 (its possible other things break that I don't even know to test).