Install models

Each framework is installed into its own conda env under envs/<env>/. Conda (or mamba) must be on PATH.

Ask your LLM

Install MACE and SevenNet and check they run on my GPU.

The agent installs each env, compiles what your GPU needs, and reports a model only once it has computed energy and forces there.

One or several models

source env.sh
./install.sh MACE                      # a framework name (MACE) or an env name (mace)
./install.sh MACE SevenNet ORB         # several at once
./install.sh --dry-run MACE            # print the plan, install nothing
./install.sh --status MACE             # ready / partial / broken / not installed

install.sh runs the same steps every time: create the env from envs/<env>.yml (or envs/<env>.build.sh where one conda solve is not enough), install catbench, prepare weights, compile what must be compiled for your GPU (NequIP with OpenEquivariance, Allegro with CuEquivariance, the D3 kernel), and mark the env ready. An interrupted install is verified and adopted or rebuilt, never duplicated.

Check that it works

python scripts/setup_verify.py MACE-MPA-0 --json

The model must compute energy and forces on the GPU; a clean exit of install.sh alone does not count.

Replay a verified build exactly

Every verified build is recorded in envs/locks/<env>.conda.txt and envs/locks/<env>.pip.txt. To install exactly that package set, with no dependency resolution:

OMM_USE_LOCK=1 ./install.sh MACE

A plain ./install.sh builds from the recipe in envs/<env>.yml (and, for a few envs, envs/<env>.build.sh) and lets pip resolve what the recipe does not pin. The recipe's pins are build inputs, and a later install step can lift one: several recipes pin setuptools below 81 so source builds still find pkg_resources, and the finished env ends with a newer one. The lock records the env as it was when it passed verification, which is why the two can differ. python3 scripts/verify_determinism.py lists every such difference, and fails if a recipe or build script installs anything without an exact version.

Use an env you already have

python3 scripts/adopt_env.py MACE ~/miniconda3/envs/MACE   # verified before it is trusted
python3 scripts/adopt_env.py --list

Gated models

UMA and eSEN weights are gated. Accept the license on the model page with your Hugging Face account, then run hf auth login before installing. See Gated models and Hugging Face token.