generative-dungeon-master

Implementation of ideas from generative agents using the power of LLM, T2S, and S2T networks.

Linux Installation

You can install the package locally with poetry. This allows fast iteration and experimentation.

1. Setup SSH for access to the repository.

2. Install pyenv - a really nice way to manage/install specific Python versions on Linux systems.

   i) Make sure you install the build dependencies for your system.

   ii) Run

   curl https://pyenv.run | bash
   

   iii) Restart your shell so the path changes take effect.

   exec $SHELL
   

   iv) Install a specific Python version (e.g. 3.8.3) with

   pyenv install 3.8.3
   

3. Install poetry for your OS.

4. Install package

   git clone git@github.com:dzermas/generative-dungeon-master.git
   cd generative-dungeon-master
   pyenv install $(cat .python-version)
   poetry install
   

5. Set up pre-commit to ensure all commits to adhere to black and PEP8 style conventions.

   poetry run pre-commit install
   

Execute locally with poetry

To see all options, simply use the CLI from the poetry environment and run:

poetry run generativedm --help

The world generation with the default parameters can be executed as:

poetry run generativedm generate-world

Docker

The default docker image supports a CPU deplyment with a lightweight python:3.8-slim-buster (less than 3GB).

If an NVidia GPU with CUDA is available, the nvidia/cuda:11.7.1-cudnn8-runtime-ubuntu20.04 is usualy a good image choice (~8GB) to start from. The Dockerfile will need to change and the lines

# Install package:
RUN poetry install
RUN pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

will have to be replaced by

# Install package:
RUN poetry install -E pytorch

Build with

docker build -t gdm-image .

from within the repo root folder.

Use the docker image with

docker run --rm -ti gdm-image poetry run generativedm generate-world

Note that since the text-generation model needs to be downloaded, using docker will have to download the model each time the image is opened. Maybe you should avoid using the --rm and have a persistent container or add a presistent volume.

LLM Models

The code can either use an OpenAI account for inference with ChatGPT4, or HuggingFace for local inference with an open network like Alpaca. The OpenAI account charges real money for inferences and it can start getting expensive especially when the code base is not efficient enough. By default, the use_openai parameter in the main.py script is set to False. The HuggingFace model is downloaded locally in the ~/.cache/hub/ folder the first time it is called by the generate() function and every subsequent use happens by loading that local model. We can experiment with other models based on the inference capabilities of the local machines. The model selection is exposed for easy experimentation.

Resources

Look into the Alpaca models, an open source small model that is supposed to rival ChatGPT3:

https://medium.com/@martin-thissen/llama-alpaca-chatgpt-on-your-local-computer-tutorial-17adda704c23

https://github.com/cocktailpeanut/dalai

HuggingFace has a lot of models we can try before we need to train and finetune anything:

https://huggingface.co/models?pipeline_tag=text-generation&sort=downloads