WAFL-llm

WAFL is built to run as a two-part system. Both can be installed on the same machine. This is the LLM side of the WAFL project.

LLM side (needs a GPU to be efficient)

This is a model server for the speech-to-text model, the LLM, the embedding system, and the text-to-speech model.

Installation

In order to quickly run the LLM side, you can use the following installation commands:

pip install wafl-llm
wafl-llm start

which will use the default models and start the server on port 8080.

The installation will require MPI and Java installed on the system. One can install both with the following commands

sudo apt install libopenmpi-dev
sudo apt install default-jdk

Configuration

A use-case specific configuration can be set by creating a config.json file in the path where wafl-llm start is executed. The file should look like this (the default configuration)

{
  "llm_model": "fractalego/wafl-phi3-mini-4k",
  "speaker_model": "microsoft/speecht5_tts",
  "whisper_model": "fractalego/personal-whisper-distilled-model",
  "sentence_embedder_models": "TaylorAI/gte-tiny",
  "entailer_model": "vectara/hallucination_evaluation_model",
  "device": "cuda",
  "quantization": false
}

The models are downloaded from the HugggingFace repository. Any other compatible model should work. The device can be set to cuda or cpu. The quantization can be set to true or false and uses the vLLM fp8 option. This will only work if your GPU supports it. Each model can be set to null to deactivate it (and save memory). This is useful if you only want to use the LLM model, for example.

Docker

Run the following

docker/build.sh
docker run -p8080:8080 --env NVIDIA_DISABLE_REQUIRE=1 --gpus all wafl-llm