Local Llama 3 Chat 🦙

A simple chat application with LLama 3 using OpenVINO Runtime for inference and transformers library for tokenization.

• Model Export
• Quickstart with Docker
• Requirements
• Getting Started
• Export from HuggingFace

Model Export

Download the INT-4 quantized Meta-Llama-3.1-8B-Instruct model already converted to the OpenVINO IR format from HuggingFace using huggingface-cli with the following command:

huggingface-cli download rajatkrishna/Meta-Llama-3.1-8b-Instruct-OpenVINO-INT4 --local-dir models/llama-3.1-instruct-8b

Quickstart with Docker

• Install docker.

• Build the docker image with the following command. The source files and model weights are pulled using git, requiring an active internet connection.

  docker build -t chat-llama .
  

• Mount the model directory and start the container using:

  docker run -v $(pwd)/models:/chat-app/models -p 5000:5000 chat-llama
  

  This should start the Flask dev server available on http://localhost:5000

Requirements

• Python 3.11

Getting Started

1. Clone the repository

   git clone https://github.com/rajatkrishna/llama3-openvino
   

2. Create a new virtual environment to avoid dependency conflicts:

   python3 -m venv create .env
   source .env/bin/activate
   

3. Install the dependencies in requirements.txt

   pip install -r requirements.txt
   

4. Start the flask server from the project root using

   python3 -m flask run
   

Export from HuggingFace

• To export the meta-llama/Meta-Llama-3-8B-Instruct model quantized to INT-8 format yourself using optimum-intel CLI, install the requirements in requirements_export.txt:

  pip install -r requirements_export.txt
  

  Then run the following from the project root:

  optimum-cli export openvino --model meta-llama/Meta-Llama-3.1-8B-Instruct --weight-format int4 models/llama-3.1-instruct-8b
  

• Alternately, use the following steps to export the INT-4 quantized model using the Python API:

  1. Import the dependencies:

     >>> from optimum.intel.openvino import OVWeightQuantizationConfig, OVModelForCausalLM
     >>> from transformers import AutoTokenizer
     

  2. Load the model using OVModelForCausalLM class. Set export=True to export the model on the fly.

     >>> export_path = "models/llama-3.1-instruct-8b"
     >>> q_config = OVWeightQuantizationConfig(bits=4, sym=True, group_size=128)
     >>> model = OVModelForCausalLM.from_pretrained(model_name, export=True, quantization_config=q_config)
     >>> model.save_pretrained(export_path)
     

  3. Now use AutoTokenizer to save the tokenizer.

     >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
     >>> tokenizer.save_pretrained(export_path)