Example of finetuning with QLoRA/QDoRA?
#2
by
asmith26
- opened
Hi there,
I'm trying to finetune this on a 16GB Colab GPU - just wondering if there are any examples of doing this (preferably with the Jax backend + QDoRA for performance)?
Thanks for any help! :)
The closest example I found is this tutorial, which demonstrates fine-tuning Gemma with LoRA/QLoRA using the JAX backend. You can follow a similar procedure to fine-tune PaliGemma with QDoRA.
Be sure to adjust the dataset and other relevant parameters to match your specific use case.
To load this specific PaliGemma model, use the following code:
preprocessor = keras_nlp.models.PaliGemmaCausalLMPreprocessor.from_preset(
"hf://google/paligemma-3b-224-mix-keras", sequence_length=256
)
paligemma_lm = keras_nlp.models.PaliGemmaCausalLM.from_preset(
"hf://google/paligemma-3b-224-mix-keras", preprocessor=preprocessor
)