[WIP] Full fine-tune and LoRA fine-tune simplifications

[Andrei-Aksionov]

Hey there 👋

Note

This is an implentation for the idea proposed in #2779

This is an implementation of my comment back from March.

The objective is to enhance the maintainability and, in certain aspects, the readability of recipes.

If we consider the differences between a full fine-tuning and LoRA variant, the primary distinctions should lie in:

1. The way a model is instantiated
2. The way a checkpoint is saved

This draft/PR proposes precisely this:

1. Utilizing the full fine-tuning recipe as a foundation
2. Overriding only the LoRA-specific methods

---

This approach offers several advantages:

1. More maintainable code: eliminating the need for identical modifications across multiple locations.
2. Up-to-date code: sometimes changes made to the full version may not have been incorporated into the LoRA variant.
3. More readable code: facilitating a clearer understanding of the unique aspects of LoRA fine-tuning (in terms of the recipe) for anyone interested in exploring it.

However, there are a few drawbacks to consider:

1. Increased readability challenges: while the single-file approach allows users to read the whole recipe as a single file, this approach requires opening two files (ideally side-by-side) to read the recipe.
2. Reduced hackability: if someone intends to use the LoRA recipe as a starting point and subsequently build upon it, the first two files must be merged. This process is straightforward, as methods need to be replaced as a whole, requiring minimal mental effort and taking a few minutes at most.

---

As with any solution, it has its own set of drawbacks. Despite these disadvantages, this approach may still be worthwhile.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/2780

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.