SFT fine-tuning reproducible recipes

[vadimkantorov]

Curious, can torchtitan consume torchtune-like recipes? Is torchtitan more scalable/performant than torchtune?

Also wondering if the two frameworks share common components / idioms?

For instance, I found that torchtune doesn't support HF tokenizers, I wonder if torchtitan is better in this regard?

• Integrate Hugging Face tokenizers with torchtune torchtune#2706

Thanks!