README.md

swarm-container

This repo builds a SwarmUI-ready container with:

• flash_attn @ 2.7.4
• sageattention @ 2.2.0
• sageattn @ 3 (compiled)
• torchaudio @ 2.9.1 (compiled)

It is built on top of the nvidia PyTorch images nvcr.io/nvidia/pytorch.

Requirements

• A Blackwell GPU
  • RTX 50-series
  • RTX Pro 6000
  • RTX Pro 5000
• Docker or Podman

Getting Started

The image is available on DockerHub, so all you need to do is have the SwarmUI repo cloned locally. Replace /path/to/SwarmUI with the path you've cloned SwarmUI at locally and run one of the following:

All model paths as default

docker run --gpus all --rm -it --shm-size=512m --name swarmui \
    -p 7801:7801 \
    -v /path/to/SwarmUI:/workspace \
    jtreminio/swarmui:latest

Then navigate to http://localhost:7801/.

Define different model and config paths

docker run --gpus all --rm -it --shm-size=512m --name swarmui \
    -p 7801:7801 \
    -v /path/to/SwarmUI:/workspace \
    -v /path/to/local/output_directory:/workspace/Output \
    -v /path/to/local/wildcard_directory:/workspace/Data/Wildcards \
    jtreminio/swarmui:latest

Then navigate to http://localhost:7801/.

Building

If you would like to build the image for yourself, simply run:

# compiles flash_attn, sageattention, torchaudio, etc
./step-1.sh
# builds the Docker image for reuse
./step-2.sh

There are two steps because docker build does not have a --gpus all option, so you cannot compile anything that requires a GPU.