Fine-tuning with TorchTune

Fine-Tuning with Llama 3

Francesca Donadoni

Curriculum Manager, DataCamp

The components of TorchTune fine-tuning

Model
- Defines the architecture and pre-trained weights to fine-tune
- Different versions and number of parameters available
Dataset
- Specifies the data used for training
Recipe
- Central configuration file combining the model, dataset, and training parameters
- Ensures consistency and reproducibility

The components of TorchTune fine-tuning

Model

!tune ls

llama3/8B_full
llama3_1/8B_full
llama3_2/1B_full ...

Dataset
- ds.save_to_disk("new_dataset")
Recipe
- custom_recipe.yaml

The components of a TorchTune recipe

General Settings and ouput directory
- Batch size, device and epochs

Model
- Specifies architecture and model configurations

Optimizer
- Includes learning rate
Dataset
- Defines preprocessing and dataset path

batch_size: 4
device: cuda
epochs: 20
output_dir: /tmp/full-llama3.2-finetune

model:
  _component_: 
      torchtune.models.llama3_2.llama3_2_1b

optimizer:
  _component_: bitsandbytes.optim.PagedAdamW8bit
  lr: 2.0e-05

dataset:
  _component_: torchtune.datasets.alpaca_dataset

Configuring TorchTune recipes

More parameters available
Configurable in Python using yaml

import yaml

config_dict = {"batch_size": 4,
               "device": "cuda",
               "model": {
                 "_component_": "torchtune.models.llama3_2.llama3_2_1b"
                },
                ...
}

yaml_file_path = "custom_recipe.yaml"
with open(yaml_file_path, "w") as yaml_file:
    yaml.dump(config_dict, yaml_file)

Running custom fine-tuning

tune run --config custom_recipe.yaml

INFO:torchtune.utils.logging:Running 
Writing logs to /tmp/full-llama3.2-finetune/log_1732815689.txt
INFO:torchtune.utils.logging:Model is initialized with precision torch.bfloat16.
INFO:torchtune.utils.logging:Tokenizer is initialized from file.
1|52|Loss: 2.3697006702423096:   0%|?                     | 52/25880

Saved logs
Successful initialization
Epoch and step count progress
Loss metrics

Let's practice!

Fine-Tuning with Llama 3