Bump matgl from 4.0.2 to 4.0.3

[dependabot]

Bumps matgl from 4.0.2 to 4.0.3.

Release notes

Sourced from matgl's releases.

v4.0.3

• **New: LAMMPS integration for TensorNet and M3GNet potentials (

Changelog

Sourced from matgl's changelog.

4.0.3

• New: LAMMPS integration for TensorNet and M3GNet potentials (#815). matgl.ext.lammps.LAMMPSMatGLModel exports a PyG Potential to a TorchScript artifact (via the new mgl create-lammps-model CLI subcommand), consumed by a pair_matgl CPU pair style and a pair_matgl/kk Kokkos GPU pair style shipped under lammps/. The export wrapper uses a kernel-composition pattern (_TensorNetKernel / _M3GNetKernel) with the strain/autograd machinery in the outer module; M3GNet required pure-tensor, script-safe ports of the three-body indexer and basis (create_line_graph_torch, _m3gnet_three_body_basis_torch). Includes drop-in CMake snippets, build instructions, parity tests, and CI jobs. Single-GPU only for Kokkos (multi-rank Kokkos + libtorch is unreliable). Fixes a ghost-row folding bug that caused a ~30-42 eV energy gap vs the ASE calculator.
• Fix: SoftExponential activation autograd correctness and NaN safety (#788). forward now selects its alpha < 0 / alpha > 0 / alpha ≈ 0 branches with torch.where instead of a Python if on the learnable alpha parameter. The old if self.alpha < 0.0 forced a host-device sync and dropped the branch from the autograd graph (so alpha was effectively trapped in its initial sign region); the alpha == 0.0 exact-float test was unreachable after the first optimizer step; and the alpha < 0 formula produced NaN/Inf for sufficiently negative inputs. The log argument and denominators are now guarded so both the activation and alpha.grad stay finite. Values in the well-defined region are unchanged.
• New: matgl.utils.MCDropoutWrapper for uncertainty-aware inference. Enables Monte Carlo Dropout (Gal & Ghahramani, 2016) on any pretrained MatGL model (CHGNet, M3GNet, TensorNet, …) without retraining: the backbone stays in eval() while only the readout dropout is sampled, and predict_uncertainty(structures, n_passes) returns per-structure (mean, std) for acquisition functions such as UCB (mean - lambda * std). See issue #800.
• Perf: backbone-once fast path for predict_uncertainty (cache_backbone=True, default). Since MC Dropout only perturbs the readout, the deterministic backbone is evaluated once and only the cheap stochastic head is replayed n_passes times (vectorised), giving an ~n_passes× speed-up (~19× at n_passes=20 on M3GNet, GPU). Numerically equivalent to the naive loop; engaged only when a probe proves the head is the model's terminal op, otherwise falls back automatically (e.g. CHGNet, which pools after dropout).
• Fix: training checkpoints loadable under torch.load(weights_only=True) (#802). ModelLightningModule / PotentialLightningModule pickled optimizer/scheduler objects (and a numpy element_refs array) into the checkpoint hyperparameters, so resuming via Trainer.fit(ckpt_path=...) broke under torch ≥ 2.6's weights_only=True default. Optimizer/scheduler are now excluded from save_hyperparameters (their state is already in optimizer_states / lr_schedulers) and element_refs is stored as a plain list.
• Fix: silent mismatch between dataset and model element lists (#819). Added a guard that reads element_types from the dataset's converter (the actual source that stamps graph.node_type) and validates it against model.element_types, catching a common fine-tuning misconfiguration that previously failed silently.
• Fix: load MatPES datasets from JSONL on Hugging Face. MGLDatasetLoader now downloads the line-delimited .jsonl files (dataset and per-element atomrefs) that materialyze/matpes moved to; monty's loadfn parses them transparently.
• New: stress warning added to JAXPESCalculator.

Commits

• a5d5afc Update version and changelog.
• 8091a95 Fix example-notebook regressions surfaced by executing all notebooks (#820)
• 61605e8 Fix for silent mismatch between dataset element list and model element list (...
• 02b415b fix: cast numpy int to Python int in compute_line_graph loop (#818)
• 632b2d4 Update uv.lock.
• e0db75c fix: load MatPES datasets from JSONL files on HF
• 451b7e5 Bug fix for split_dataset.
• 53780e7 Updated all notebooks.
• 2891142 fix(examples): repair training notebooks and move datasets to HF Hub
• ec72545 Update tutorials.
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)