Load_in_8bit vs. loading 8-bit quantized model

[Meshwa]

Use LoRA it works with as little as a T4 15 gb, only fine-tuning the small model and then merging it with the big boi