LLMFlex

LLMFlex is a python package that allows python developers to work with different large language models (LLM) and do prompt engineering with a simple interface. It favours free and local resources instead of using paid APIs to develop truly local and private AI-powered solutions.

It provides classes to load LLM models, embedding models, and vector databases to create LLM powered applications with your own prompt engineering and RAG techniques. With a one-liner command, you can load a chatbot interface to chat with the LLM or serve a model as OpenAI API as well.

Installing LLMFlex

Creating a virtual environment before installing the package is highly recommended. Also make sure you have installed Pytorch and llama-cpp-python with the correct installation method according to your hardware configuration before installing LLMFlex. Please visit the links provided for the respective packages for more detailed installation guides.

After you have done the above steps, you can easily install llmflex with pip.

pip install llmflex

You can also install ExLlamaV2, AutoAWQ, and AutoGPTQ if you have CUDA devices. Please visit the links provided for the respective packages for more detailed installation guides.

Features

1. Multilple LLMs with different genration configurations from one model

Unlike Langchain, you can create multiple LLMs with different temperature, max new tokens, stop words etc. with the same underlying model without loading the model several times using the LlmFactory class. This can be useful when you create your own agent with different LLM tasks which requires different configurations.

2. Langchain compatibility with enhanced performances

All the LLMs created with LlmFactory are langchain compatible, and can be seamlessly integrated in your existing Langchain code. All the LLM classes are re-implementations of some langchain LLM classes which support more efficient streaming and stop words management, all with a unified interface.

3. Multiple model formats support

Multiple model formats are all supported, and the loading process are all handled in the LlmFactory class, so it is just plug and play. Supported formats:

• PyTorch (using transformers)
• GGUF (using llama-cpp-python)
• AWQ (using transformers)
• GPTQ (using transformers)
• OpenAI API (Work with any local servers that serve models with OpenAI API format)
• EXL2 (using exllamav2)

4. Embedding Toolkits

Bundled classes for using embedding models which contains the embedding model and a tokens-count-based text splitter using the embedding model.

5. Vector database

Utilising Embedding toolkits and FAISS, a VectorDatabase class can allow you to store and search texts for your RAG tasks.

6. Chat memories

Chat memory classes for storing chat memory on disk.

1. BaseChatMemory
   Memory class without using any embedding models or vector databases.

2. LongShortTermChatMemory
   Memory class using an underlying VectorDatabase to maintain long term memory along with the most recent memory.

7. Prompt template

A PromptTemplate class is implemented to format your prompt with different prompt formats for models from different sources. Some presets like Llama2, ChatML, Vicuna, and more are already implemented, but you can alway add your own prompt format template.

8. Custom tools

A base class BaseTool for creating llm powered tools. A WebSearchTool powered by DuckDuckGo is implemented as an example.

9. Chatbot frontend interface

If you simply want to play with a model, there's a gradio frontend chatbot that allows you to chat with a model with different generation configurations. You can switch between chat histories and prompt format, and you can set your system prompt and other model text generation sampling configurations in the gradio webapp.

Using LLMFlex

1. Create LLMs

This is how you can start with any text generation model on HuggingFace with your machine.

from llmflex import LlmFactory

# Load the model from Huggingface
model = LlmFactory("TheBloke/OpenHermes-2.5-Mistral-7B-GGUF")

# Create a llm
llm = model(temperature=0.7, max_new_tokens=512, stop=['```'])

# Use the LLM for your task
prompt = """Instruction:
Write a python script to load a Pandas Dataframe from the csv file 'test.csv'

Solution:
```python
"""
script = llm.invoke(prompt)
print(script)

# Or if you prefer to generate the output with token streamming.
for token in llm.stream(prompt):
    print(token, end="")

2. Load embeddings toolkit and create vector database

To load an embedding model and use a vector database:

from llmflex.Embeddings import HuggingfaceEmbeddingsToolkit
from llmflex.VectorDBs import FaissVectorDatabase

# Loading the embedding model toolkit
embeddings = HuggingfaceEmbeddingsToolkit(model_id="thenlper/gte-large")

# Create a vector database
food = ["Apple", "Banana", "Pork"]
vectordb = FaissVectorDatabase.from_texts(index=food, embeddings=embeddings)

# Do semantic search on the vector database
print(vectordb.search("Beef"))

3. Use memory and prompt template to create a chatbot

Use the Memory classes and the PromptTemplate class to create a chatbot.

from llmflex.Memory import LongShortTermChatMemory, create_long_short_prompt
from llmflex.Prompts import PromptTemplate

# Create the memory with the embeddings toolkit created earlier
memory = LongShortTermChatMemory(title="My first conversation with OpenHermes", embeddings=embeddings, from_exist=False)

# Get a prompt template
prompt_template = PromptTemplate.from_preset("ChatML")

# Create an llm from the model factory
chat_llm = model(max_new_tokens=512, stop=prompt_template.stop + ["###"])

# Create the prompt for the llm
user_input = "What is the meaning of life?"
prompt = create_long_short_prompt(
    user_input=user_input,
    prompt_template=prompt_template,
    llm=chat_llm,
    memory=memory
)

# Generating the response from the model
output = ""
for token in chat_llm.stream(prompt):
    output += token
    print(token, end="")

# Save the interaction between you and the chatbot
memory.save_interaction(user_input=user_input, assistant_output=output)

4. Use tools

A WebSearchTool class is implemented as an example to build a tool with LLMFlex. The tool is using DuckDuckGo by default. Here is how you can use it:

from llmflex.Tools import WebSearchTool

# Create a web search tool with the embeddings toolkit created earlier
tool = WebSearchTool(embeddings=embeddings, verbose=True)

# Run the tool with the previsous chat memory for a more coherent response
tool_input = "Can you give me some views of the meaning of life from some famous philosophers?"
output_generator = tool.run(tool_input, 
    llm=chat_llm, 
    history=memory.get_recent_memory(last_k=2), 
    prompt_template=prompt_template,
    stream=True)

tool_output = ""
for token in output_generator:
    tool_output += token
    print(token, end="")

# And save the response to the memory
memory = memory.save_interaction(user_input=tool_input, assistant_output=tool_output)

5. Chat with the model in a Streamlit web app

If you just want a GUI to start chatting with your LLM model with both long term and short term memory, type this command in the terminal:

llmflex interface --model_id TheBloke/OpenHermes-2.5-Mistral-7B-GGUF --embeddings thenlper/gte-small

If you want to chat with the web search tool:

# Use a model with longer context for web search
llmflex interface --model_id TheBloke/OpenHermes-2.5-Mistral-7B-16k-GGUF --embeddings thenlper/gte-small --web_search --extras "model_file='openhermes-2.5-mistral-7b-16k.Q6_K.gguf', context_length=16384"

You will see a streamlit frontend, use it to chat with the LLM model.

6. Serve an OpenAI API with a GGUF model

To serve a GGUF model with OpenAI API:

llmflex serve --model_id TheBloke/OpenHermes-2.5-Mistral-7B-GGUF --model_file openhermes-2.5-mistral-7b.Q6_K.gguf --context_size 4096

Documentations

Python documentation for all the classes, methods, and functions is provided in the ./docs directory in this repository.

License

This project is licensed under the terms of the MIT license.