Running Llama 2 and other Open-Source LLMs on CPU Inference Locally for Document Q&A

Clearly explained guide for running quantized open-source LLM applications on CPUs using LLama 2, C Transformers, GGML, and LangChain

Step-by-step guide on TowardsDataScience: https://towardsdatascience.com/running-llama-2-on-cpu-inference-for-document-q-a-3d636037a3d8

Context

• Third-party commercial large language model (LLM) providers like OpenAI's GPT4 have democratized LLM use via simple API calls.
• However, there are instances where teams would require self-managed or private model deployment for reasons like data privacy and residency rules.
• The proliferation of open-source LLMs has opened up a vast range of options for us, thus reducing our reliance on these third-party providers. 
• When we host open-source LLMs locally on-premise or in the cloud, the dedicated compute capacity becomes a key issue. While GPU instances may seem the obvious choice, the costs can easily skyrocket beyond budget.
• In this project, we will discover how to run quantized versions of open-source LLMs on local CPU inference for document question-and-answer (Q&A).
  
  

Quickstart

• Ensure you have downloaded the GGML binary file from https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML and placed it into the models/ folder
• To start parsing user queries into the application, launch the terminal from the project directory and run the following command: poetry run python main.py "<user query>"
• For example, poetry run python main.py "What is the minimum guarantee payable by Adidas?"
• Note: Omit the prepended poetry run if you are NOT using Poetry
  
  

Tools

• LangChain: Framework for developing applications powered by language models
• C Transformers: Python bindings for the Transformer models implemented in C/C++ using GGML library
• FAISS: Open-source library for efficient similarity search and clustering of dense vectors.
• Sentence-Transformers (all-MiniLM-L6-v2): Open-source pre-trained transformer model for embedding text to a 384-dimensional dense vector space for tasks like clustering or semantic search.
• Llama-2-7B-Chat: Open-source fine-tuned Llama 2 model designed for chat dialogue. Leverages publicly available instruction datasets and over 1 million human annotations.
• Poetry: Tool for dependency management and Python packaging

Files and Content

• /assets: Images relevant to the project
• /config: Configuration files for LLM application
• /data: Dataset used for this project (i.e., Manchester United FC 2022 Annual Report - 177-page PDF document)
• /models: Binary file of GGML quantized LLM model (i.e., Llama-2-7B-Chat)
• /src: Python codes of key components of LLM application, namely llm.py, utils.py, and prompts.py
• /vectorstore: FAISS vector store for documents
• db_build.py: Python script to ingest dataset and generate FAISS vector store
• main.py: Main Python script to launch the application and to pass user query via command line
• pyproject.toml: TOML file to specify which versions of the dependencies used (Poetry)
• requirements.txt: List of Python dependencies (and version)

References

• https://github.com/marella/ctransformers
• https://huggingface.co/TheBloke
• https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML
• https://python.langchain.com/en/latest/integrations/ctransformers.html
• https://python.langchain.com/en/latest/modules/models/llms/integrations/ctransformers.html
• https://python.langchain.com/docs/ecosystem/integrations/ctransformers
• https://ggml.ai
• https://github.com/rustformers/llm/blob/main/crates/ggml/README.md
• https://www.mdpi.com/2189676