Create interface for accessing the training code

[RobotSail]

This PR turns this into a library.

View the following PR for an example: instructlab/instructlab#1329

The main idea is that we provide a function like:

def train_torchrun(torchargs: TorchrunArguments, training_args: FullTrainingArguments):
  pass

Each class TorchrunArguments and FullTrainingArguments provides training-specific arguments. And then from another library, you would simply provide these arguments to the train_torchrun function.

For any other training method that we define, we could provide a similar interface depending on which arguments are needed.

We separate the arguments here because TorchrunArguments are the ones passed to torchrun and then the full training arguments are the ones that we actually train with. It's not crucial that these are different, but it makes our lives a lot easier from a maintenance standpoint.

[RobotSail]

instructlab/eval#1

[russellb]

I haven't reviewed the code part, just pyproject/requirements/namespace package portions. That all looks good, though I gave some suggested edits for some minor cleanup. That includes dropping Python 3.9 since I think we landed on 3.10 as the minimum.

[nathan-weinberg]

Just focused on the Python packaging stuff, will leave the training code itself to y'all for now 😄

[Maxusmusti]

Note: Current requirements will overwrite existing nvidia pytorch installs. Need to ensure that if those exist, we are not installing our own torch.

[JamesKunstle]

two thoughts, one being a minor nit-pick, but in general this is very good, I'll check that it runs.

[RobotSail]

Thank you for the review @JamesKunstle, I've created an issue about your quantization comment here: #29

[cdoern]

initial comments. Want to make sure this is aligned with CLI impl.

[russellb]

Can you squash the commit history here before merging?

[JamesKunstle]

v good thank you for all your work

[russellb]

biggestion question here is about the use of class variables.

In terms of PR structure -- it would have been nice to have the PR that's doing the library setup setup from all the API introduction stuff.

[Maxusmusti]

LGTM, tested full training (no lora/offload) with both base scripts and updated interface, both functioning as expected

[russellb]

just to be super clear:

1. squash the first two commits and fix up the commit message
2. create a follow-up issue for cleaning up requirements.txt to specify versions for all dependencies + making sure the versions are aligned with instructlab/instructlab if it's a shared dependency

[russellb]

2. create a follow-up issue for cleaning up requirements.txt to specify versions for all dependencies + making sure the versions are aligned with instructlab/instructlab if it's a shared dependency

We talked about this, so I know your intent was to create this issue, but I can't find it anywhere. I went ahead and filed this one: #34