Blame for README.md

---
license: mit
---

# 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.

| Env name | Stack | Key packages |
|---|---|---|
| `grace` | GRACE (tensorpotential) | tensorflow |
| `nequip` | NequIP / Allegro | nequip, torch 2.12 (cpu) |
| `esen` | eSEN (fairchem) | fairchem-core, torch 2.4.1 (cpu), torch_scatter/sparse |
| `tace2` | TACE | tace (git pin), torch 2.13 (cpu) |
| `mbd` | SevenNet and general MLIP tooling | sevenn, torch 2.8 (cpu), ase, pymatgen |

## Quick start

```bash
# 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).

```bash
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/`.