The cost of full fine tuning
Full fine tuning updates every weight, so each task needs a complete model copy and heavy optimizer memory. Parameter efficient fine tuning, often called PEFT, instead freezes most of the model and trains only a small number of new or selected parameters.
The shared idea
- The large pretrained weights stay frozen.
- A tiny set of trainable parameters is added or chosen.
- Only those parameters receive gradients and are stored per task.
This cuts memory for optimizer states and makes each task a small adapter file rather than a full checkpoint.
The structure
Why it matters
PEFT methods such as adapters, prefix tuning, and low rank updates can reach accuracy close to full fine tuning while training a fraction of the parameters. They make it practical to serve many tasks from one frozen backbone by swapping small modules, and they lower the hardware bar for adapting large models.
Key idea
Parameter efficient fine tuning freezes the backbone and trains only a small set of parameters, approaching full fine tuning quality while slashing memory and per task storage.