mlip-stack-docker

MLIP Stack Docker (CPU)

Ready-to-run Docker image bundling five machine-learning interatomic potential (MLIP) stacks in isolated conda environments. Built for CPU-only machines (no CUDA required). All environments use Python 3.11.

Quick start

# download and load (~7.6 GB compressed, ~28 GB on disk)
wget https://huggingface.co/datasets/Seanmonami/mlip-stack-docker/resolve/main/mlip-router-stack_2026-08-21.tar.gz
docker load < mlip-router-stack_2026-08-21.tar.gz

# interactive shell
docker run --rm -it mlip-router-stack:2026-08-21

# run something in a specific env
docker run --rm -v $PWD:/work -w /work mlip-router-stack:2026-08-21 \
  micromamba run -n mbd python your_script.py

Inside the container, switch environments with micromamba run -n <env> ... or micromamba activate <env>.

Rebuilding from source

docker_build/ contains the Dockerfile and per-env requirement files (pip freezes normalized to CPU wheels). env_snapshots/ contains the original conda env export YAMLs and raw pip freezes captured from the source machine (Ubuntu 24.04, 2026-08-21).

cd docker_build && bash build.sh

Notes

• Model checkpoints are NOT included; download them via each package's own tooling (e.g. sevenn, fairchem, GRACE model registry).
• CUDA wheels were replaced with CPU wheels (nvidia-* and triton dropped). For GPU use, rebuild with the original pins in env_snapshots/.