README.md

Fine-tune Phi3 using Olive

In this example you'll use Olive to:

1. Fine-tune a LoRA adapter to classify phrases into Sad, Joy, Fear, Surprise.
2. Merge the adapter weights into the base model.
3. Optimize and Quantize the model into int4.

We'll also show you how to inference the fine-tuned model using the ONNX Runtime (ORT) Generate API.

⚠️ For Fine-tuning, you'll need to have a suitable GPU available - for example, an A10, V100, A100.

💾 Install

Create a new Python virtual environment (for example, using conda):

conda create -n olive-ai python=3.11
conda activate olive-ai

Next, install the Olive and the dependencies for a fine-tuning workflow:

cd Phi-3CookBook/code/04.Finetuning/olive-ort-example
pip install olive-ai[gpu]
pip install -r requirements.txt

🧪 Fine-tune Phi3 using Olive

The Olive configuration file contains a workflow with the following passes:

Phi3 -> LoRA -> MergeAdapterWeights -> ModelBuilder

At a high-level, this workflow will:

1. Fine-tune Phi3 (for 150 steps, which you can modify) using the dataset/data-classification.json data.
2. Merge the LoRA adapter weights into the base model. This will give you a single model artifact in the ONNX format.
3. Model Builder will optimize the model for the ONNX runtime and quantize the model into int4.

To execute the workflow, run:

olive run --config phrase-classification.json

When Olive has completed, you're optimized int4 fine-tuned Phi3 model is available in: code/04.Finetuning/olive-ort-example/models/lora-merge-mb/gpu-cuda_model.

🧑‍💻 Integrate fine-tuned Phi3 into your application

To run the app:

python app/app.py --phrase "cricket is a wonderful sport!" --model-path models/lora-merge-mb/gpu-cuda_model

This response should be a single word classification of the phrase (Sad/Joy/Fear/Surprise).