PEFT
Parameter-Efficient Fine-Tuning (PEFT) adapts a pretrained model by training only a small number of extra or selected parameters while freezing the rest, drastically reducing memory and storage costs.
What is PEFT?
Parameter-Efficient Fine-Tuning (PEFT) adapts a pretrained model by training only a small number of extra or selected parameters while freezing the rest, drastically reducing memory and storage costs.
Parameter-Efficient Fine-Tuning (PEFT) adapts a pretrained model by training only a small number of extra or selected parameters while freezing the rest, drastically reducing memory and storage costs.
Where is it used?
The Hugging Face `peft` library implements LoRA, QLoRA, prefix tuning, and adapters; models like Llama-3 and Mistral are commonly PEFT-tuned for domain adaptation on single GPUs.
How to build it
Wrap a model with `peft.get_peft_model(model, LoraConfig(task_type="CAUSAL_LM"))`, print `trainer.model.print_trainable_parameters()` to see <1% trainable params, then train with `Trainer`.