Skip to content

Latest commit

 

History

History
255 lines (174 loc) · 8.48 KB

research.md

File metadata and controls

255 lines (174 loc) · 8.48 KB
Raw

Research to TinyLLM

The following documents my research journey to building my TinyLLM, a locally run LLM and Chatbot. I have also recorded two video classes to document some of my learning:

Hardware Specs

My research began trying to prove that a LLM could run on consumer grade equipment. The reference base hardware:

  • CPU: Intel(R) Core(TM) i5-6500 CPU @ 3.20GHz
  • Memory: 8GB DDR4 (more recommended)
  • Disk: 128G SSD Boot, 4TB HDD
  • GPU: NVIDIA GeForce GTX 1060 6GB

Suggested:

  • GPU: NVIDIA GeForce RTX 3090 24GB (with CUDA Version: 12.2)

Setup

# Clone the llama.cpp project
git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp

# Linux - Build for Nvidia GPU using CMake
mkdir build
cd build
cmake .. -DLLAMA_CUBLAS=ON   # Omit -DLLAMA_CUBLAS=ON for CPU only
cmake --build . --config Release

Using LLaMA-2 or Mistral 7B - 5-bit Quantized Models

# Download GGUF models from HuggingFace.
cd models

# Meta LLaMA-2 7B GGUF Q-5bit model Q5_K_M
wget https://huggingface.co/TheBloke/Llama-2-7b-Chat-GGUF/resolve/main/llama-2-7b-chat.Q5_K_M.gguf

# Mistral 7B GGUF Q-5bit model Q5_K_M
wget https://huggingface.co/TheBloke/Mistral-7B-Claude-Chat-GGUF/resolve/main/mistral-7b-claude-chat.Q5_K_M.gguf

cd ..

# Run interactive chat.
./build/bin/main -m models/llama-2-7b-chat.Q5_K_M.gguf \
    -t 4 \
    --color \
    -c 4096 \
    --temp 0.7 \
    --gpu-layers 32 \
    -n -1 \
    -i -ins 

Where:
    -m models/llama-2-7b-chat.Q5_K_M.gguf   # The model
    -t 4                                    # change to match number of CPU cores
    -c 4096                                 # context length
    --temp 0.7                              # randomness 
    --gpu-layers 32                         # number of layers to offload to GPU - remove if cpu only
    -n -1 --color                           # options 
    -i -ins                                 # interactive mode and instruction

See https://github.com/ggerganov/llama.cpp/blob/master/examples/main/README.md for more details on inference parameters.

Example chat using llama.cpp interactive mode:

./main -t 4 -m models/llama-2-7b-chat.Q5_K_M.gguf \
    --color -c 4096 --temp 0.7 --gpu-layers 32 -n -1 -i -ins
> Pick a color
Green

> What is a llama?
A llama is a large, domesticated mammal that is native to South America. It has long, shaggy fur and distinctive ears with a distinctive curled-over shape. Llamas are used for their wool and as pack animals in remote areas where cars cannot reach. They are also known for their calm and gentle nature.

> Write a haiku
Llama in the sun
Gentle eyes, shaggy coat
Soft as a cloud

This example model was run on a Ubuntu Linux host with an Intel i5-6500 CPU @ 3.20GHz, 8GB RAM and an Nvidia GTX 1060 GPU with 6GB VRAM.

LLaMA-2-Linux_NVIDIA-GTX-1060-small.mov

Python API

The models built or downloaded here can be used by the LLaMa-cpp-python project.

# Linux OS - Build and Install with Nvidia GPU support
CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python==0.2.27

This will also build llama.cpp but includes the python bindings. Next, if you downloaded the Llama-2 LLM model above, you can test it using this python script:

from llama_cpp import Llama

# Load model - use gpu for 32 of 35 NN layers to keep it within the 6GB VRAM limit
llm = Llama(model_path="models/llama-2-7b-chat.Q5_K_M.gguf", n_gpu_layers=32)

# Ask a question
question = "Name the planets in the solar system?"
print(f"Asking: {question}...")
output = llm(f"Q: {question} A: ", 
    max_tokens=64, stop=["Q:", "\n"], echo=True)

# Print answer
print("\nResponse:")
print(output['choices'][0]['text'])

OpenAI API Compatible Server

The llama-cpp-python library has a built in OpenAI API compatible server. This can be used to host your model locally and use OpenAI API tools against your self-hosted LLM.

Manual Setup

# Install Server that uses OpenAI API
CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python[server]==0.2.7

# Run the API Server
python3 -m llama_cpp.server \
    --model ./models/llama-2-7b-chat.Q5_K_M.gguf \
    --host localhost \
    --n_gpu_layers 32 

# It will listen on port 8000

Run as a Service

See instructions here: https://github.com/jasonacox/TinyLLM/tree/main/llmserver

CLI Chat using API

See the example chat.py CLI Chatbot script that connects to this server and hosts an interactive session with the LLM.

The example chat.py Features:

  • Use of OpenAI API (could be used to connect to the OpenAI service if you have a key)
  • Works with local hosted OpenAI compatible llama-cpp-python[server]
  • Retains conversational context for LLM
  • Uses response stream to render LLM chunks instead of waiting for full response

Example Test Run (./chat.py):

ChatBot - Greetings! My name is Jarvis. Enter an empty line to quit chat.

> What is your name?

Jarvis> Jarvis.

> What is today's date?

Jarvis> Today's date is September 10, 2023.

> What day of the week is it?

Jarvis> It is Sunday.

> Answer this riddle: Ram's mom has three children, Reshma, Raja and a third one. What is the name of the third child?

Jarvis> The answer to the riddle is "Ram."

> Pick a color.

Jarvis> Jarvis will choose blue.

> Now write a poem about that color.

Jarvis> Here is a short poem about the color blue:
Blue, the hue of the sky so high
A symbol of hope, and a sight to the eye
Soothing and calm, yet bold and bright
The color of serenity, and pure delight.

> What time is it?

Jarvis> The current time is 10:45 AM.

> Thank you very much!

Jarvis> You're welcome! Is there anything else I can assist you with?

Web Based Chatbot

The above CLI chat was converted to a web based python flask app in the chatbot folder. To run the web server you can use the docker container as specified in the chatbot folder or do this manually:

# Install required packages
pip install openai flask flask-socketio bs4
cd chatbot

# Run the chatbot web server - change the base URL to be where you host your llmserver
OPENAI_API_BASE="http://localhost:8000/v1" python3 server.py

Open http://127.0.0.1:5000 - Example session:

image

Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation (RAG) is an architecture that combines the strengths of retrieval-based and generation-based language models. The basic idea is to use a retrieval model to generate high-quality text, and then augment this text with additional information from a generative model. This allows RAG models to generate more accurate and informative text than either retrieval or generation models alone.

See the rag folder for information about setting up the vector database, embedding your documents and integrating with the LLM (e.g. llama).

Train

The llama.cpp project includes a train-text-from-scratch tool. Use -h to see the options or an example below.

# Create a text file to use for training
mkdir models/jason
curl -s https://github.com/jasonacox/ProtosAI/files/11715802/input.txt > models/jason/jason.txt

# Run the training
./train-text-from-scratch \
        --vocab-model models/ggml-vocab.bin \
        --checkpoint-in models/jason/jason.chk.bin \
        --checkpoint-out models/jason/jason.chk.bin \
        --model-out models/jason/jason.bin \
        --train-data models/jason/jason.txt \
        --ctx 32 --embd 256 --head 8 --layer 16 \
        -t 4 -b 32 --seed 42 --adam-iter 16 \
        --use-flash --print-details-interval 0 --predict 64 \
        -n 1 # adjust this for the number of iterations to run

References

Additional Tools