PET — Garden's Almanac

Checkpoint	Params	Polaris	Sophia	Perlmutter	Delta	Frontier	Della
pet-oam-xl	730M	●	●	—	○	—	○
pet-omatpes-l	255M	●	●	—	●	—	●

● verified, last 30 days ○ installed, not recently verified — not installed

No cluster has installed any checkpoint of this model yet.

Running this model

1# from a job or interactive session on a supported cluster:

2from rootstock import RootstockCalculator

4with RootstockCalculator(

5 cluster=YOUR_CLUSTER_ID, # eg, "polaris", "sophia"

6 checkpoint="pet-oam-xl",

7 device="cuda",

8) as calc:

9 # model is now running in subprocess on compute node

10 atoms.calc = calc

11 atoms.get_potential_energy()

Environments

Rootstock runs each model family inside an isolated Python environment defined by a single file. This environment file includes the specific dependencies needed, plus a setup() function that loads the model and returns an ASE calculator. These files are usually almost identical for a given model family, but because of cluster-specific quirks (eg, an old CUDA driver) the dependencies and setup code can vary a bit.

polaris-pet_env.py

1# /// script

2# requires-python = ">=3.11,<3.15"

3# dependencies = [

4# "upet>=0.2.6",

5# "ase>=3.22",

6# # upet pulls nvalchemi-toolkit-ops unpinned; 0.4+ needs torch>=2.8 at

7# # runtime but only declares the constraint on its extras, so the

8# # resolver won't catch it (same trap as the tensornet env).

9# "torch>=2.8,<2.14",

10# ]

11#

12# [tool.uv.sources]

13# torch = { index = "pytorch-cu128" }

14#

15# [[tool.uv.index]]

16# name = "pytorch-cu128"

17# url = "https://download.pytorch.org/whl/cu128"

18# explicit = true

19# ///

20"""PET env — Polaris variant of pet.py (UPET foundation checkpoints).

22Identical to pet.py except torch is pinned to the cu128 index: Polaris's

23driver stack tops out at CUDA 12.8, and the default PyPI torch wheel is

24built against a newer CUDA, so metatomic's pick_device finds no usable

25cuda at setup() (smoke-test 2026-08-04). Sophia keeps the unrestricted env.

27The upstream string encodes model@version; versions are pinned rather than

28"latest" so rebuilds serve the same weights. pet-omatpes-l is trained at the

29r2SCAN level of theory — its energies are not comparable to the PBE-level

30pet-oam models.

31"""

33CHECKPOINTS = {

34 "pet-oam-xl": "pet-oam-xl@1.0.0",

35 "pet-omatpes-l": "pet-omatpes-l@0.1.0",

36}

38CLUSTERS = ["polaris"]

41def setup(checkpoint: str, device: str = "cuda"):

42 from huggingface_hub import hf_hub_download

43 from upet.calculator import UPETCalculator

45 # metatomic-ase 0.1.2's nvalchemi fast path passes a float max_neighbors

46 # (len * max(128, cutoff**3)) into torch.full — TypeError for cutoffs > ~5 Å.

47 # Force the vesin fallback; the flag is read at call time.

48 import metatomic_ase._neighbors as _mta_n

49 _mta_n.HAS_NVALCHEMIOPS = False

51 # Passing model=/version= makes UPETCalculator resolve the name by listing

52 # the hub repo — an uncached API call that fails on workers, which run

53 # with HF_HUB_OFFLINE=1 (and on any node without internet). Fetch the

54 # pinned file ourselves — a cache hit needs no network even offline — and

55 # hand it over as checkpoint_path, which skips the resolve entirely.

56 model, version = CHECKPOINTS[checkpoint].split("@", 1)

57 path = hf_hub_download(

58 repo_id="lab-cosmo/upet",

59 filename=f"{model}-v{version}.ckpt",

60 subfolder="models",

61 )

62 return UPETCalculator(checkpoint_path=path, device=device)

Built on Polaris: 2026-08-05

Couldn't load the current environments from Rootstock.

References