Important

NeMo 2.0 is an experimental feature and currently released in the dev container only: nvcr.io/nvidia/nemo:dev. Please refer to NeMo 2.0 overview for information on getting started.

Parameter-Efficient Fine-Tuning (PEFT)

The following instructions explain how to fine-tune and evaluate a T5 model using PEFT.

Fine-Tune with PEFT

This section explains how to fine-tune with the NeMo Launcher and configure training.

Fine-Tune with NeMo Launcher

  1. To run finetuning with PEFT, first update conf/config.yaml:

defaults:
  - peft: t5/squad

stages:
  - peft

  1. Then, execute the launcher pipeline: python3 main.py

Configure Training

You can find default configurations for PEFT with squad in conf/peft/t5/squad.yaml. Fine-tuning configuration is divided into four sections run, trainer, exp_manger and model.

  1. Configure the training metadata:

run:
  name: peft_t5_220m
  time_limit: "04:00:00"
  dependency: "singleton"
  convert_name: convert_nemo
  model_train_name: t5_220m
  convert_dir: ${base_results_dir}/${peft.run.model_train_name}/${peft.run.convert_name}
  task_name: "squad"
  results_dir: ${base_results_dir}/${.model_train_name}/peft_${.task_name}

  1. Set the number of nodes and devices for fine-tuning:

trainer:
  num_nodes: 1
  devices: 8

  1. Set the path to the checkpoint from which to restore the model:

model:
  restore_from_path: ${peft.run.convert_dir}/results/megatron_t5.nemo

restore_from_path sets the path to the .nemo checkpoint to run fine-tuning.

  1. Set the PEFT-specific configuration:

model:
    peft:
        peft_scheme: "lora"

peft_scheme sets the fine-tuning scheme to be used. Supported schemes include: lora, adapter, ia3, ptuning.

Evaluate with PEFT

Evaluate with NeMo Launcher

  1. To run evaluating PEFT, first update conf/config.yaml:

defaults:
  - evaluation: peft_t5/squad

stages:
  - evaluation

  1. Then, execute the launcher pipeline: python3 main.py.

Configure Evaluation

You can find the default configurations for PEFT with squad in conf/evaluation/peft_t5/squad.yaml.

  1. Set the path to checkpoint to restore model from.

model:
  restore_from_path: ${peft.run.convert_dir}/results/megatron_t5.nemo

restore_from_path sets the path to the .nemo checkpoint to run fine-tuning.

  1. Set PEFT specific configruation and path to the PEFT weights file to restore from.

model:
    peft:
        peft_scheme: "lora"
        restore_from_path: ${base_results_dir}/${.model_train_name}/peft_${.task_name}

peft_scheme sets the fine-tuning scheme to be used. restore_from_path sets the path to the .nemo PEFT weights checkpoint.