Docker usage¶

KonfAI ships Docker assets in the repository under docker/.

The Docker image is designed for CLI-oriented workflows executed from a mounted workspace.

Official image¶

The repository documentation states that the official image is published as:

docker pull vboussot/konfai

Build locally¶

docker build -f docker/Dockerfile -t konfai .

Useful build arguments documented in docker/README.md:

KONFAI_PYPI_VERSION
KONFAI_EXTRAS
TORCH_INDEX_URL

Run the low-level CLI¶

Training:

docker run --rm -it \
  --gpus all \
  -v "$(pwd):/workspace" \
  -w /workspace \
  vboussot/konfai TRAIN --gpu 0 -c examples/Synthesis/Config.yml

Prediction:

docker run --rm -it \
  --gpus all \
  -v "$(pwd):/workspace" \
  -w /workspace \
  vboussot/konfai PREDICTION --models checkpoint.pt --gpu 0 -c examples/Synthesis/Prediction.yml

Evaluation:

docker run --rm -it \
  -v "$(pwd):/workspace" \
  -w /workspace \
  vboussot/konfai EVALUATION -c examples/Synthesis/Evaluation.yml

Run the app layer¶

docker run --rm -it \
  --gpus all \
  -v "$(pwd):/workspace" \
  -w /workspace \
  vboussot/konfai konfai-apps infer my_app -i input.mha -o ./Output

Notes¶

The container defaults to konfai --help.
GPU visibility still depends on the host runtime and NVIDIA container support.
For host troubleshooting, see docker/README.md.