Tiny llm Finetuner for Intel dGPUs

Finetuning openLLAMA on Intel discrete GPUS

A finetuner^{1 2} for LLMs on Intel XPU devices, with which you could finetune the openLLaMA-3b model to sound like your favorite book.

Setup and activate conda env

conda env create -f env.yml
conda activate pyt_llm_xpu

Warning: OncePyTorch and intel extension for PyTorch is already setup, then install peft without dependencies as peft requires PyTorch 2.0(not supported yet on Intel XPU devices.)

Generate data

Fetch a book from guttenberg (default: pride and prejudice) and generate the dataset.

python fetch_data.py

Finetune

python finetune.py --input_data ./book_data.json --batch_size=64 --micro_batch_size=16 --num_steps=300

Inference

For inference, you can either provide a input prompt, or the model will take a default prompt

Without user provided prompt

python inference.py --infer

Using your own prompt for inference

python inference.py --infer --prompt "my prompt"

Benchmark Inference

python inference.py --bench

1: adapted from: source ↩
2: adapted from: source ↩