Runs an AIDungeon2 fork in Docker on AMD ROCm hardware.
The Clover-Edition fork of AIDungeon2 and the KoboldAI project use PyTorch, which supports AMD ROCm capable GPUs in addition to the more widely used Nvidia CUDA GPUs.
This project builds a Docker container based on rocm/pytorch, containing everything to play the game on an AMD GPU - tested using a Vega 56 card. Using the 16-bit model, response times are down to a few seconds at most, compared to half a minute or more on CPU-only runs. The 8GB VRAM is more than enough to support the 16-bit model. This image uses a fork (finetuneanon/transformers) of the Huggingface transformers, which enables GPT-Neo to run in 8GB of VRAM.
Install Docker and the AMD ROCm kernel module on your host (see https://rocmdocs.amd.com/en/latest/Installation_Guide/Installation-Guide.html).
Run ./build.sh to build the Docker image. The models and saved games will be stored in the persistent-data directory, which is mounted into the Docker container.
Test your installation using ./run.sh /app/test/test.sh. It should output detailed info about your GPU and compute capabilities. It should end like this:
Testing ROCm-CUDA on pyTorch
CUDA(hip) is available: True
CUDA(hip) device_count: 1
Device name: Device 687f
To run the game, execute ./run.sh . by default, this launches an KoboldAI instance.
To monitor the GPU usage on your host, execute watch /opt/rocm/bin/rocm-smi .