LLM finetune

This codebase shows how to quickly set up an efficient finetune using Unsloth for large language models on a SLURM-based system.

Install

Install the following code via ssh. Ideally, on a GPU node which you will run the finetune on such that it can pick up the right GPU architecture. For example, I will finetune on a H100 to I did the installation also on a H100.

module load 2023
module load Python/3.11.3-GCCcore-12.3.0 CUDA/12.1.1
python -m venv venv
source venv/bin/activate
pip install -U pip
pip install "unsloth[cu121-ampere-torch240] @ git+https://github.com/unslothai/unsloth.git"

For troubleshooting unsloth, please refer to here

Usage

Specify within your job script what kind of GPU, for how long and the hyperparameters of your finetuning script. Here is also where you change the dataset and model to your own experiments. Then start training with:

sbatch finetune_llama.job

Experimental setup

Dataset

Dataset Name	Total Samples	Average Tokenized Length	Max Token Length
Slim Orca	517,982	408	8212

Model

Dataset Name	Model parameters	Context length	QLoRA
Llama 3.1 8B 4-bit	8M	131k	~0.5% trainable parameters

Results

python finetune_unsloth.py \
                          --pretrained_model_name_or_path unsloth/Meta-Llama-3.1-8B-bnb-4bit \
                          --data_dir Open-Orca/SlimOrca \
                          --output_dir /scratch-shared/$USER/finetune_results/ \
                          --max_seq_length 8192 \
                          --per_device_train 8 \
                          --per_device_eval 8 \
                          --num_train_epochs 1 \
                          --optim adamw_8bit \
                          --bf16 \
                          --gradient_accumulation_steps 4 \
                          --packing \
                          --logging_steps 100 \

On a single H100 GPU took 8.5 hours