Full Fine-Tuning

About

The d9d.peft.full_tune package brings standard fine-tuning into the PEFT workflow. It does not change the model architecture. It unfreezes all parameters of the modules whose names fully match a regular expression, e.g. normalization layers or a specific head. Full fine-tuning is most useful together with other PEFT methods through Method Stacking. For example, you can apply LoRA to attention layers and fully fine-tune the normalization layers.

Usage

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

import re
from d9d.peft import inject_peft_and_freeze
from d9d.peft.full_tune import FullTune, FullTuneConfig

model = ...  # Your model.

method = FullTune(FullTuneConfig(module_name_pattern=re.compile(r".*norm.*")))

# Freezes every parameter except those of modules named like "...norm...".
# Full fine-tuning keeps the state keys, so the returned mapper is empty.
inject_peft_and_freeze(method, model)

API Reference