yeah, i was wondering well, what about fine tuning a already trained model. still need a really nice home lab
didn't know about this, but what's the point. just max vram and just load a model someone else did
Quote:
|
Parameter-Efficient Fine-Tuning (PEFT): This is a game-changer for consumer hardware. Techniques like QLoRA significantly reduce VRAM requirements by only fine-tuning a small portion of the model. Using QLoRA, a 7B model can often be fine-tuned on a single GPU with less than 10-12 GB of VRAM
|
7B model?, who gives a shit.