Skip to content

Text generation in Python, as easy as possible

License

Notifications You must be signed in to change notification settings

ddh0/easy-llama

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

easy-llama

PyPI PyPI - Python Version PyPI - License

Text generation in Python, as easy as possible

>>> import easy_llama as ez
>>> Mixtral = ez.Model('Mixtral-8x7B-v0.1-q8_0.gguf')
>>> Mixtral.generate('The sky is blue, and')
' the grass is green. It seems like the most natural thing in the world to most of us. However, have you ever stopped to think that the color of these things is actually a perception of our brain?'
>>> 

easy-llama's purpose is to make use of on-device large language models (LLMs) as easily as possible. It is a layer of abstraction over llama-cpp-python, which itself provides the Python bindings for the underlying llama.cpp project.

Documentation

Documentation is available in DOCS.md.

Features

  • Automatic arbitrary context length extension
    • Just specify your desired context length, and easy-llama will adjust the necessary parameters accordingly
    • A warning will be displayed if the chosen context length is likely to cause a loss of quality
  • Terminal-based interactive chat
    • Thread.interact()
    • Optional text streaming
    • Different colored text to differentiate user / bot
    • Some basic commands accessible by typing ! and pressing ENTER:
      • reroll - Re-roll/swipe last response
      • remove - Remove last message
      • reset - Reset the Thread to its original state without re-loading model
      • clear - Clear the screen
      • sampler - View and modify sampler settings on-the-fly
  • Programmatic multi-turn interaction
    • Thread.send(prompt) -> response
    • Both your message and bot's message are added to Thread
  • Several common prompt formats built-in
    • accessible under ez.formats
    • Stanford Alpaca, Mistral Instruct, Mistral Instruct Safe, ChatML, Llama2Chat, Llama3, Command-R, Vicuna LMSYS, Vicuna Common, Dolphin, Guanaco, & more
    • Easily extend, duplicate and modify built-in formats
    • ez.formats.wrap(prompt) - Wrap a given prompt in any prompt format for single-turn completion
  • Message-based context length handling
  • Retrieve sorted list of candidates for the most likely next token (i.e. logits)
  • Support all models supported by llama-cpp-python

Examples

Programmatic chat

>>> import easy_llama as ez
>>> Llama3 = ez.Model('Meta-Llama-3-8B-Instruct-q8_0.gguf')
>>> Thread = ez.Thread(Llama3, ez.formats.llama3, ez.samplers.TikTokenSampling)
>>> Thread.send('tell me a fun fact about lions')
'Roar! Hi there!\n\nHere\'s a fun fact about lions: Did you know that lions are social butterflies... er, I mean, social big cats? They live in prides, which are typically made up of several females, their cubs, and one or more males. But here\'s the cool part: lionesses often give each other "grooming massages" by licking and nuzzling each other to strengthen their bonds and show affection! Who knew lions could be so... pampered?\n\nWant to learn more fun facts about lions?'
>>> Thread.send('now tell me a joke about them')
'Here\'s one that\'s the "mane" event:\n\nWhy did the lion go to the dentist?\n\nBecause it had a ROAR-ing toothache! (get it?)\n\nHope that made you roar with laughter! Do you want another one?'
>>> 

Interactive chat

>>> import easy_llama as ez
>>> Llama3 = ez.Model('Meta-Llama-3-8B-Instruct-q8_0.gguf')
>>> Thread = ez.Thread(Llama3, ez.formats.llama3, ez.samplers.TikTokenSampling)
>>> Thread.interact()
Click to expand interactive chat example

In practice, the prompt and the response are differentiated by color for improved readability.

  > What is HTML?

Hello there! I'm Llama 3, here to help you with any questions you may have.

HTML stands for HyperText Markup Language. It's the standard markup language used to create web pages. In simple terms, HTML is a way to add structure and content to a webpage, making it look and behave as intended.

HTML consists of a series of elements, represented by tags (<>). These tags are used to define different parts of a webpage, such as headings, paragraphs, images, links, forms, tables, and more. For example, the `<p>` tag is used to define a paragraph of text, while the `<img>` tag is used to add an image to a webpage.

HTML documents typically consist of:
  
1. Headings: <h1>, <h2>, <h3>, etc.
2. Paragraphs: <p>
3. Links: <a>
4. Images: <img>
5. Lists: <ul>, <ol>, <li>
6. Tables: <table>, <tr>, <td>
7. Forms: <form>, <input>, <textarea>, <select>

HTML is used in conjunction with other technologies, such as CSS (Cascading Style Sheets) for styling and JavaScript for adding interactivity to a webpage.

HTML5 is the latest version of HTML, which introduced new features like semantic elements, improved forms, and multimedia support. It's widely supported by modern web browsers and is used to build most websites on the internet today.

I hope that helps you understand what HTML is! Do you have any other questions about it?

  > How is that different from python?

Python and HTML are two completely different programming languages, with different purposes and uses. Here's a brief comparison:

**Purpose:**

* HTML (HyperText Markup Language) is used for creating the structure and content of web pages. Its primary purpose is to define the layout, format, and organization of information on the web.
* Python is a high-level, interpreted programming language that can be used for a wide range of purposes, including:
    + Web development (e.g., building web applications, scraping websites)
    + Data analysis and science
    + Machine learning and artificial intelligence
    + Automation and scripting
    + Game development

**Syntax:**

* HTML uses markup tags to define the structure and content of a webpage. It's primarily used for defining the layout and organization of information.
* Python uses indentation (spaces or tabs) to denote block-level syntax, and is typically written in a text editor or IDE.

**Programming paradigm:**

* HTML is a declarative language, meaning you describe what you want to see on the page, rather than how to achieve it. It's used for defining the structure of a webpage.
* Python is an imperative language, meaning you provide step-by-step instructions on how to solve a problem or perform a task.

**Output:**

* HTML produces an output that is rendered by web browsers, resulting in a visual representation of the content.
* Python can produce various outputs, such as:
    + Text files
    + Images
    + Audio files
    + Web pages (using frameworks like Flask or Django)
    + Data analysis results

In summary, HTML is used for creating the structure and content of web pages, while Python is a programming language that can be used for a wide range of purposes, including web development, data analysis, machine learning, and more. While they're both important tools in web development, they serve different purposes and are used in different ways.

Do you have any other questions about HTML or Python?

  > no that's it thanks

You're welcome! It was a pleasure chatting with you and helping you understand the basics of HTML. If you have any more questions or need further clarification, feel free to reach out anytime.  

Have a great day and happy coding (or web-browsing)!

  > 

Installing with pip

In most cases, the best way to install easy-llama is through pip.

You will need cmake to install easy-llama (except for CPU only). It is probably available in your preferred package manager, such as apt, brew, yum, etc. Or you can install it from source.

Select your backend from the list below to see your installation instructions. If you run into issues with the installation, please see the llama-cpp-python installation instructions for a more detailed guide. If you're still having trouble, feel free to open an issue.

CPU only
pip uninstall llama-cpp-python -y
pip install --no-cache-dir llama-cpp-python
pip install --upgrade --no-cache-dir easy-llama
CUDA (for NVIDIA)
pip uninstall llama-cpp-python -y
CMAKE_ARGS="-DGGML_CUDA=1" pip install --no-cache-dir llama-cpp-python 
pip install --upgrade --no-cache-dir easy-llama
Metal (for Apple Silicon)
pip uninstall llama-cpp-python -y
CMAKE_ARGS="-DGGML_METAL=1" pip install --no-cache-dir llama-cpp-python 
pip install --upgrade --no-cache-dir easy-llama
ROCm (for AMD)
pip uninstall llama-cpp-python -y
CMAKE_ARGS="-DGGML_HIPBLAS=1" pip install --no-cache-dir llama-cpp-python 
pip install --upgrade --no-cache-dir easy-llama
Vulkan
pip uninstall llama-cpp-python -y
CMAKE_ARGS="-DGGML_VULKAN=1" pip install --no-cache-dir llama-cpp-python 
pip install --upgrade --no-cache-dir easy-llama
OpenBLAS
pip uninstall llama-cpp-python -y
CMAKE_ARGS="-DGGML_BLAS=1 -DGGML_BLAS_VENDOR=OpenBLAS" pip install --no-cache-dir llama-cpp-python 
pip install --upgrade --no-cache-dir easy-llama
SYCL
pip uninstall llama-cpp-python -y
source /opt/intel/oneapi/setvars.sh
CMAKE_ARGS="-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx" pip install --no-cache-dir llama-cpp-python 
pip install --upgrade --no-cache-dir easy-llama

Installing from source

Installation from source is only necessary if you rely on some functionality from llama.cpp that is not yet supported by llama-cpp-python. In most cases, you should prefer installing with pip.

You will need cmake. It is probably available in your preferred package manager, such as apt, brew, yum, etc. Or you can install it from source.

Note

You will need to modify the CMAKE_ARGS variable according to your backend. The arguments shown below are for CUDA support. If you're not using CUDA, select your backend above to see the correct CMAKE_ARGS.

To install easy-llama from source, copy and paste the following commands into your terminal:

pip uninstall easy-llama llama-cpp-python -y
rm -rf ./easy-llama
rm -rf ./llama-cpp-python
git clone "https://github.com/abetlen/llama-cpp-python"
cd ./llama-cpp-python/vendor/
rm -rf ./llama.cpp
git clone "https://github.com/ggerganov/llama.cpp"
cd -
git clone "https://github.com/ddh0/easy-llama"
CMAKE_ARGS="-DGGML_CUDA=1" pip install -e ./llama-cpp-python
pip install -e ./easy-llama

Note that installations from source are bleeding-edge, and as such are inherently unstable. You may run into unexpected issues during installation or inference, which is why installing from pip is recommended in most cases.

Acknowledgments

easy-llama stands on the shoulders of giants. Thank you to Andrei Betlen for llama-cpp-python, and to Georgi Gerganov for llama.cpp and GGML. Thank you to all who have made contributions to these projects.

DISCLAIMER

All language models tend to produce writing that is factually inaccurate, stereotypically biased, and fundamentally disconnected from reality.

LICENSE

The Unlicense

easy-llama is free and unencumbered software released into the public domain