Deathsaurus

Discord bot to interface with Transformer models.

Prerequisites

You need an NVIDIA GPU to run the models. Requires Python 3.10+ and the appropriate CUDA libraries installed on your system (including cuDNN). Tested in WSL2 (Ubuntu) on Windows 10.

Jax can't be installed directly by pip in all cases -- see instructions here.

You may need to set the following environment variable:

# You may get an error about compilation parallelism if this isn't set
export XLA_FLAGS="--xla_gpu_force_compilation_parallelism=1"

# Make sure your CUDA libraries are on LD_LIBRARY_PATH --
# this is the default install location for WSL2
export CUDA_HOME=/usr/local/cuda
export PATH="$CUDA_HOME/bin:$PATH"
export LD_LIBRARY_PATH="$CUDA_HOME/lib64:$LD_LIBRARY_PATH"

Usage

Build the requirements:

make build

Run a REPL for bot commands:

# Text generation
make run-local-text

# Image generation
make run-local-image

Export some needed environment variables and run the bot connected to Discord:

export DISCORD_BOT_TOKEN="<your token>"
export DISCORD_BOT_GUILD="<your server name>"
export DISCORD_BOT_CHANNEL="<channel where bot should post>"
export DISCORD_BOT_HOF_CHANNEL="<channel where bot should make hall-of-fame posts>"

# Text generation
make run-discord-text

# Image generation
make run-discord-image

Commands

• !deathsaurus_help: Display usage
• !deathsaurus_ping: Ping the bot to check if it's online
• !deathsaurus_gen <text>: Generate text from the given seed
• !deathsaurus_hof: Print instructions on using the Hall of Fame

You can also @ the bot to run commands:

• @<bot_name> help
• @<bot_name> ping
• @<bot_name> gen
• @<bot_name> hof

Parameter Guidelines

Text Generation

For the GPT2-Large model:

• temperature = 0.1: Boring, very realistic text
• temperature = 0.7: Mostly realistic text
• temperature = 0.99: Coherent yet very whacky text
• temperature > 5: Incoherent mass of words