peterspackman/mlip-gguf

MLIP GGUFs

GGUF-format conversions of the uPET family of machine-learning interatomic potentials, for use with mlip.cpp and mlip.js.

Source

Checkpoints converted from lab-cosmo/upet (BSD-3-Clause). See the LICENSE file in this repo.

Usage

from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="peterspackman/mlip-gguf", filename="pet-mad-s.gguf")

Or in the browser via mlip.js:

const buf = await fetch(
  "https://huggingface.co/peterspackman/mlip-gguf/resolve/main/pet-mad-s.gguf"
).then(r => r.arrayBuffer())
const model = await Model.loadFromBuffer(buf)

Files

• pet-mad-s.gguf (95.4 MB)
• pet-mad-xs.gguf (16.3 MB)
• pet-oam-l.gguf (721.9 MB)
• pet-omad-s.gguf (95.4 MB)
• pet-omad-xs.gguf (16.3 MB)
• pet-omat-s.gguf (95.4 MB)
• pet-omat-xs.gguf (16.3 MB)
• pet-spice-s.gguf (53.4 MB)

Conversion

These files are produced by scripts/convert_models.py in the mlip.cpp repo, which wraps scripts/export_pytorch/export_pet_gguf.py (an exact torch.export of the PyTorch forward + backward graph into GGUF tensors + a graph interpreter preamble).