Open LLM WebUI

This repository contains a web application designed to execute relatively compact, locally-operated Large Language Models (LLMs).

Installation

Please follow these steps to install the software:

• Create a new conda environment:

conda create -n ollm python=3.10
conda activate ollm

• Clone the software repository:

git clone https://github.com/Uminosachi/open-llm-webui.git
cd open-llm-webui

Python Package Installation

General Instructions

• Install the necessary Python packages by executing:

  pip install -r requirements.txt

Installation for Flash Attention (Optional)

• To enable Flash Attention in some models, if CUDA is available, install Flash Attention:

  pip install packaging ninja
  pip install flash-attn --no-build-isolation

Platform-Specific Instructions

• For Windows (with CUDA support):

  • Install Visual Studio:

    • ⚠️ Important: Make sure to select Desktop development with C++ during the installation process.

  • Copy MSBuild extensions for CUDA as an administrator (adjust the CUDA version v12.1 as needed):

    xcopy /e "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.1\extras\visual_studio_integration\MSBuildExtensions" "C:\Program Files\Microsoft Visual Studio\2022\Community\MSBuild\Microsoft\VC\v170\BuildCustomizations"

  • Configure the required environment variables for the build (adjust the CUDA version as necessary):

    set PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.1\bin;%PATH%
    "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat"
    set FORCE_CMAKE=1
    set CMAKE_ARGS="-DGGML_CUDA=ON -DCMAKE_CXX_FLAGS=/utf-8 -DCMAKE_C_FLAGS=/utf-8"
    set CMAKE_BUILD_PARALLEL_LEVEL=16

  • Install the necessary Python packages (this process may take some time):

    pip install ninja cmake scikit-build-core[pyproject]
    pip install --force-reinstall --no-cache-dir llama-cpp-python
    pip install -r requirements.txt

• For Linux (with CUDA support):

  • Configure the required environment variables for the build (if not already set):

    export PATH=/usr/local/cuda/bin:${PATH}
    export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}

  • Install the OpenMP libraries used for the build:

    sudo apt-get update
    sudo apt-get install libgomp1 libomp-dev

  • Install the necessary Python packages:

    pip install ninja cmake scikit-build-core[pyproject]
    export CMAKE_ARGS="-DGGML_CUDA=ON"
    pip install --force-reinstall --no-cache-dir llama-cpp-python
    pip install -r requirements.txt

• For Mac OS (without CUDA support):

  • Install the necessary Python packages:

    BUILD_CUDA_EXT=0 pip install -r requirements.txt

  • Rebuild the bitsandbytes package with the CPU option.

    pip uninstall bitsandbytes
    git clone https://github.com/TimDettmers/bitsandbytes.git
    cd bitsandbytes
    cmake -DCOMPUTE_BACKEND=cpu -S .
    make
    pip install .

  • Known Issue: Running the LLaVA model on Mac results in an error.

Running the application

python ollm_app.py

• Open http://127.0.0.1:7860/ in your browser.

Downloading the Model

To download the model:

• Launch this application.
• Click on the "Download model" button next to the LLM model ID.
• Wait for the download to complete.

📜 Model List (transformers)

Provider	Model Names
Microsoft	Phi-3-mini-4k-instruct
Google	gemma-2-9b-it, gemma-1.1-2b-it, gemma-1.1-7b-it
NVIDIA	Llama3-ChatQA-1.5-8B
Qwen	Qwen2-7B-Instruct
Mistral AI	Mistral-7B-Instruct-v0.3
Rakuten	RakutenAI-7B-chat, RakutenAI-7B-instruct
rinna	youri-7b-chat
TheBloke	Llama-2-7b-Chat-GPTQ, Kunoichi-7B-GPTQ

• 📋 Note: By adding the repository paths of models to model_manager/add_tfs_models.txt, they will be included in the list of Model IDs and displayed in the UI.
• 🔍 Note: The downloaded model file will be stored in the .cache/huggingface/hub directory of your home directory.

Access and Download Gemma and Llama Models

• Before downloading any models, ensure that you have obtained the necessary access rights through Hugging Face. Please visit the following pages to request access:
  • Llama 3 model by Meta
  • Llama 2 model by Meta
  • Gemma 1.1 model by Google

Login to Hugging Face

• Before downloading any models, please log in via the command line using:

  huggingface-cli login

🦙 Model List (llama.cpp)

Provider	Model Names
Microsoft	Phi-3-mini-4k-instruct-q4.gguf, Phi-3-mini-4k-instruct-fp16.gguf
TheBloke	llama-2-7b-chat.Q4_K_M.gguf
QuantFactory	Meta-Llama-3-8B-Instruct.Q4_K_M.gguf

Using any GGUF file

• 🔍 File Placement: Place files with the .gguf extension in the models directory within the open-llm-webui folder. These files will then appear in the model list on the llama.cpp tab of the web UI and can be used accordingly.
• 📝 Metadata Usage: If the metadata of a GGUF model includes tokenizer.chat_template, this template will be used to create the prompts.

🖼️ Model List (Multimodal LLaVA)

Provider	Model Names
Microsoft	Phi-3-vision-128k-instruct
llava-hf	llava-v1.6-mistral-7b-hf, llava-v1.6-vicuna-7b-hf, llava-1.5-7b-hf
tinyllava	TinyLLaVA-Phi-2-SigLIP-3.1B
openbmb	MiniCPM-V-2_6-int4, MiniCPM-V-2_6, MiniCPM-Llama3-V-2_5-int4, MiniCPM-Llama3-V-2_5
SakanaAI	EvoVLM-JP-v1-7B

Access and Download Llama 3 Models

• Before downloading any models, ensure that you have obtained the necessary access rights through Hugging Face. Please visit the following pages to request access:
  • Llama 3 model by Meta

Usage

• Enter your message into the "Input text" box. Adjust the slider for "Max new tokens" as needed.
• Under "Advanced options" adjust the settings for "Temperature", "Top k", "Top p", and "Repetition Penalty" as needed.
• If replacing the system message of the prompt, under "Advanced options" enable the checkbox and enter text.
• Press "Enter" on your keyboard or click the "Generate" button.
  • ⚠️ Note: If the cloud-based model has been updated, it may be downloaded upon execution.
• If you click the "Clear chat" button, the chat history will be cleared.

transformers tab

• By enabling the CPU execution checkbox, the model will use the argument device_map="cpu".
• Some of the transformers models are loaded with the following 4-bit or 8-bit settings using the bitsandbytes package.

llama.cpp tab

• Use the radio buttons in the Default chat template to select the template that will be used if the GGUF model lacks a chat_template.

LLaVA tab

• You can upload an image to the LLaVA Image area of this tab and input a prompt related to the image.
• Some of the LLaVA models are loaded with the following 4-bit or 8-bit settings using the bitsandbytes package.

options

• When you enable the Translate (ja->en/en->ja) checkbox:
  • Any input in Japanese will be automatically translated to English, and responses in English will be automatically translated back into Japanese.
  • ⚠️ Note: Downloading the translation model for the first time may take some time.

Model Credit

Developer	Model	License
Meta	Llama-3.2	Llama 3.2 Community License
Meta	Llama-3.1	Llama 3.1 Community License
Meta	Llama-3	Llama 3 Community License
Meta	Llama-2	Llama 2 Community License
Microsoft	Phi-3.5, Phi-3	The MIT License
Google	Gemma	Gemma Terms of Use
NVIDIA	Llama3-ChatQA	Llama 3 Community License
Alibaba Group	Qwen2-7B-Instruct	Apache License 2.0
Mistral AI	Mistral-7B-Instruct	Apache License 2.0
Rakuten	RakutenAI	Apache License 2.0
rinna	Youri	Llama 2 Community License
Sanji Watsuki	Kunoichi-7B	CC-BY-NC-4.0
Hugging Face	llava-v1.6-mistral-7b-hf	Apache License 2.0
Hugging Face	llava-v1.6-vicuna-7b-hf, llava-1.5-7b-hf	Llama 2 Community License
TinyLLaVA	TinyLLaVA-Phi-2-SigLIP-3.1B	Apache License 2.0
OpenBMB	MiniCPM	MiniCPM Model License
Sakana AI	EvoVLM-JP-v1-7B	Apache License 2.0