aurora_configs: native FairChem XPU via collaborator fork; add esen + allscaip

[OwenPriceSkelly]

What

Builds on #217/#219 to widen FairChem coverage on Aurora:

• aurora_configs/uma.py: replace the MLIPPredictUnit._setup_device monkeypatch with fairchem-core installed from the xpu-support branch of abagusetty/fairchem, which adds native Intel-GPU support — device="xpu" accepted through get_predict_unit/FAIRChemCalculator.from_model_checkpoint, a common/device_utils.py abstraction layer, torch.xpu seeding/cache management, and XCCL (oneCCL) collectives. The #219 conventions are preserved (mandatory task head, VERIFY_KWARGS, FP64 inference settings, **kwargs).
• aurora_configs/{esen,allscaip}.py (new): the other two fairchem-core v2 families, mirroring their nvidia_configs counterparts plus the XPU wheel index, the same FP64 settings, and an xpu _fairchem_device() shim — the fork normalizes indexed device strings to the bare type and resolves the actual GPU from torch.xpu.current_device(), so "xpu:N" has to travel through torch.xpu.set_device, same as the CUDA configs.
• docs/environments.md: refresh the aurora examples line (monkeypatch mention is stale).

Out of scope: the fairchem-core 1.x OC20 stack (escn, equiformer) — no XPU path exists (CUDA-only pyg wheels), and the fork doesn't touch v1.

Why the fork

PyPI fairchem-core (through 2.22) asserts device in {"cpu", "cuda"} deep in the predict unit; the monkeypatch worked around the assert but left cache management, seeding, and collectives CUDA-hardwired. The fork (3 commits ahead of upstream main, head fc18598 as of 2026-08-28) does this properly and is intended for upstreaming. The git dependency is routed via [tool.uv.sources] with subdirectory = "packages/fairchem-core"; env builds lock it to a concrete SHA at install time. Once XPU support merges upstream, the source override can be deleted and the dep returned to a PyPI floor.

Validation

• All three configs resolve with uv lock --script (the same resolution path rootstock install uses): fairchem-core 2.22.1.dev11+gfc18598, torch 2.13.0+xpu, triton-xpu 3.7.2. The hatch-vcs dynamic version works from the git+subdirectory install.
• Checkpoint ids verified against the fork's pretrained_models.json (includes uma-s-1p2p1).
• parse_checkpoints_dict / parse_clusters_list / declares_setup_from_path pass on all aurora configs.
• Not yet run on Intel hardware — these are experimental; first runtime validation will follow the intel workshop loop / Aurora smoke test.

🤖 Generated with Claude Code