Can you help the efforts to integrate the fine-tuning of gemma-2 it in Torchtune ?

#5
by almugabo - opened

with their multilingual capabilities gemma-2 models are a good candidates to serve as base of models in low resources languages.
There are good frameworks for fine-tuning gemma-2 but they either

  • don't offer full fine tuning (e.g. Unsloth)
  • or are resources optimized (e.g. Axolotl )
  • or are not very widely used in the community because they are tailored to gemma-2 (like the fine-tuning approach gemma-2 offered using Keras).

There is a need for a way to full-finetune gemma-2 using existing fine-tuning frameworks which work with different LLM and are resources optimized. A good candidate for that is Torchtune.
Torchtune already works well with gemma-1 but not yet with gemma-2 .

Could you help the efforts to integrate the fine-tuning of gemma-2 it in Torchtune ?

Here an issue which has been opened on that :

https://github.com/pytorch/torchtune/issues/1813

Sign up or log in to comment