- Add UMAStack thin wrapper around fairchem.core eSCNMDBackbone,
  with lazy import + clear error if fairchem-core is missing.
- Wire uma_* config keys through create_model and update_config
  (sphere_channels<-hidden_dim, lmax<-max_ell, etc.); uma_hidden_channels
  defaults to hidden_dim. Clamp uma_mmax<=max_ell to avoid zero-channel
  SO(2) convs. Robust cell/pbc handling for batched PyG graphs.
- Add fairchem-core to requirements-optional.txt.
- Add UMA to CI parametrization in tests/test_graphs.py and
  tests/test_graphs_graphattr.py (basic + equivariant suites).
- Add UMA to LennardJones training CI in
  tests/test_forces_equivariant_training.py and to the energy-only
  equivariance benchmark in tests/test_forces_equivariant.py.
- Apply black==21.5b1 formatting.

Adds the standard 10-line BSD-3-clause header to every Python file
that lacked one. The year reflects the file's last git-commit year
(falls back to mtime). Vendored FairChem trees under
examples/*/fairchem/core/** are intentionally skipped; they retain
Meta's MIT header.

Introduce a 'uma_variant' config key selecting the UMA capacity tier:
- 'S' (default): single dense eSCNMDBackbone.
- 'M' / 'L': Mixture-of-Linear-Experts (MoLE) routing via
  eSCNMDMoeBackbone, with progressively larger routed-expert counts
  (defaults M=8, L=32, overridable via 'uma_num_experts').

MoLE routing coefficients are derived from the per-system charge/spin
(and optionally atomic-composition) embeddings, so no HydraGNN-side
dataset labelling is required. New config keys: uma_variant,
uma_num_experts, uma_moe_dropout, uma_use_composition_embedding, wired
through create_model and defaulted in config_utils.

Verified: all three variants run a forward pass; the default 'S' path
still passes the UMA CI (tests/test_graphs.py -k UMA, 3 passed).

Full rewrite of the AllScAIP HydraGNN wrapper with a redesigned config
surface and completed conditioning/force-training support:

- Charge/spin conditioning: per-graph charge/spin read from data.charge
  / data.spin when present (defaulting to neutral / singlet), fed into
  the backbone's ChgSpinEmbedding.
- Dataset conditioning: new DatasetEmbedding in the AllScAIP InputBlock,
  threaded through GraphAttentionData (custom_types) and populated in
  data_preprocess. Enabled when a non-empty dataset list is configured;
  routes per-graph dataset ids into the node embeddings.
- Force training: autograd-based forces stay invariant; the soft-kNN
  radius graph keeps energy differentiable w.r.t. positions (verified
  gradient flow to data.pos).
- Config surface redesign: allscaip_* keys wired through create_model
  and defaulted in config_utils.

Verified: backbone forward with/without dataset routing and force-grad
flow to positions; AllScAIP CI passes (tests/test_graphs.py -k
AllScAIP, 2 passed).

- examples/LennardJones/LJ_UMA.json: runnable UMA (S variant) config for
  the LennardJones energy+forces example driver (--inputfile LJ_UMA.json).
- examples/LennardJones/LJ_AllScAIP.json: runnable AllScAIP config for the
  same driver.
- tests/test_example_configs_uma_allscaip.py: lightweight smoke tests that
  build a model from each shipped example config via update_config +
  create_model_config on a tiny in-memory dataset, plus coverage that the
  UMA config constructs for every S / M / L capacity tier.

All 5 smoke tests pass.

- setup.py: reformat the package_data literal to satisfy black==21.5b1
  (the version pinned in requirements-dev.txt and .pre-commit-config).
- pyproject.toml: add [tool.black] force-exclude for the verbatim vendored
  UMA mirror (hydragnn/utils/model/uma/_vendored). force-exclude is honoured
  by both CI's 'black .' and the pre-commit hook (which passes explicit
  paths), keeping the upstream-faithful mirror out of reformatting.

Verified: 'black --check .' reports all files unchanged.

The vendored FairChem UMA tree transitively imports fairchem-core's own
third-party deps (omegaconf, hydra, torchtnt, monty, ray) which are not part
of HydraGNN's core requirements. Importing eSCNMDBackbone at UMAStack module
level made 'import hydragnn' -- and therefore all CI test collection -- fail
with ModuleNotFoundError.

- UMAStack: defer the vendored import to _load_uma_backbones(), invoked only
  when a UMA model is constructed; raise a clear ImportError pointing at
  requirements-optional when the deps are absent.
- tests/uma_optional.py: uma_available() helper; UMA parametrizations now
  skip (not error) when the optional deps are missing.
- requirements-optional.txt: document the UMA backbone deps.

AllScAIP is unaffected (only needs torch/e3nn).

Investigation showed the vendored UMA backbone only reaches two lightweight
deps at import/construct/forward time -- omegaconf and monty. The heavy
fairchem-core deps (torchtnt/ray/wandb via mlip_unit; hydra via models/base)
are used solely in checkpoint/inference code paths that HydraGNN never runs;
they were merely dragged in by module-level imports.

- Vendored escn_md.py / base.py: defer the OutputSpec/Task (mlip_unit) and
  hydra imports to their single use sites. tools/vendor_uma.py now applies
  these as reproducible post-vendor PATCHES so re-vendoring keeps them.
- requirements-base.txt: add omegaconf + monty (UMA now works with core deps).
- requirements-optional.txt: drop the UMA dep block.
- UMAStack: keep the lazy backbone import (fast 'import hydragnn') with an
  updated error message pointing at requirements-base.
- tests: UMA is now a plain parametrization like MACE; removed the
  skipif scaffolding and tests/uma_optional.py.

Verified with ray/wandb/torchtnt/hydra fully blocked: import, construct,
forward, and a full forward+backward+optimizer step all succeed. 427 tests
collect clean; UMA train + smoke tests pass; black 21.5b1 clean.

Adding omegaconf==2.3.0 to requirements-base broke CI: it pulls the
sdist-only antlr4-python3-runtime, which fails to build under CI's
--no-build-isolation. Instead of shipping extra deps, guard the two
lightweight import-time deps in the vendored FairChem UMA backbone:

* escn_md.py: omegaconf DictConfig/ListConfig -> fall back to () (they
  are only used in isinstance() checks, which handle empty tuples).
* radius_graph_pbc_nvidia.py, atomic_data.py: monty.dev.requires ->
  faithful fallback decorator (the guarded functions are never called
  by HydraGNN).

UMA now needs zero dependencies beyond HydraGNN's core requirements.
Reverts the requirements-base additions and records the three guards
as reproducible PATCHES in tools/vendor_uma.py.

Verified: UMA unit tests pass and 427 tests collect with ray, wandb,
torchtnt, hydra, omegaconf and monty all blocked at interpreter startup.

Replaces the try/except import fallbacks in the vendored FairChem UMA
backbone with plain imports, and instead declares omegaconf and monty
in requirements-optional.txt. CI now installs them in a dedicated step
that keeps build isolation enabled, so omegaconf's sdist-only
antlr4-python3-runtime dependency can build its own backend (this is
what previously broke the --no-build-isolation base install).

Changes:
* escn_md.py / radius_graph_pbc_nvidia.py / atomic_data.py: restore the
  verbatim upstream imports of omegaconf / monty.dev.requires.
* tools/vendor_uma.py: drop the three guard PATCHES (keep only the
  mlip_unit / hydra deferrals); update header comment.
* requirements-optional.txt: add omegaconf and monty.
* .github/workflows/CI.yml: install omegaconf + monty with build
  isolation before the torch/pyg/deepspeed steps.
* UMAStack.py: point the lazy-import error message and NOTE at
  requirements-optional.txt.

Verified: black clean, 427 tests collect, vendored UMA imports resolve.

omegaconf/monty differ from the rest of requirements-optional: they are
required whenever UMA runs (and UMA is tested in CI), so they get a
dedicated build-isolated CI install step. Add a section header
explaining this so the separation is not lost in the flat list.

Instead of hardcoding omegaconf/monty in the CI workflow, put them in a
dedicated, informatively-named requirements file for model-specific
backbone dependencies. The file header documents that it must be
installed WITH pip build isolation (omegaconf pulls the sdist-only
antlr4-python3-runtime).

* requirements-specific-models.txt: new file (omegaconf, monty).
* CI.yml: install '-r requirements-specific-models.txt' (build isolated)
  and add it to the pip cache key.
* install_dependencies.sh: install it too (build isolated).
* requirements-optional.txt: drop the UMA block (now lives in the new file).
* UMAStack.py / tools/vendor_uma.py: point references at the new file.

All ORNL BSD headers already matched the hydragnn/models/Base.py format
exactly; only the year varied inconsistently (2021/2024/2025/2026).
Set each file's copyright year to that file's git last-commit year so
the header format stays identical to Base.py and only the year differs
per file. 124 files updated, each changed by exactly one line.