Testing out more efficient collate function + Lora

[sachinruk]

Hello 👋! I made a more efficient collate function and use Lora to try and fine tune the same dataset that you have in your notebook. However, I keep running out of GPU memory. Wondering if you have any tips, or if you are simply doing this on a A100, could you help me test this out 🙏. Happy to pay for any compute used, just don't know how to set up that side of things.

https://www.kaggle.com/code/sachin/moondream2-finetuning