Bump matgl from 4.0.3 to 4.1.0

[dependabot]

Bumps matgl from 4.0.3 to 4.1.0.

Release notes

Sourced from matgl's releases.

v4.1.0

• Disk-backed PyG training for datasets larger than memory. write_mgl_shards streams records into versioned, transactional CPU shards; MGLDiskDataset loads one shard per worker on demand; and ShardBatchSampler keeps batches shard-local while assigning disjoint shards and equal step counts to distributed ranks. MGLDataLoader now selects this path automatically for disk datasets without changing its existing in-memory behavior. MGLDataModule provides Lightning fit/validate/test/predict loaders with deterministic epoch shuffling, multiworker support, automatic MatGL collation, and optional on-the-fly graph conversion. Rebuild failures retain the prior committed manifest and clean up incomplete shards, while successful rebuilds remove superseded shards. A runnable QET notebook demonstrates sharded dataset creation, Lightning training, and per-graph QEq charge conservation.
• Improved PyG command-line workflows. mgl train now trains or fine-tunes interatomic potentials from local MatPES-shaped JSON/JSONL or Extended XYZ data through MGLDatasetLoader and MGLPotentialTrainer; mgl evaluate evaluates a saved potential with force/stress autograd enabled. Extended XYZ input supports periodic structures and nonperiodic molecules, standard or user-selected label keys, and TensorNet scratch training. The JSON loader accepts both current record-oriented MatPES files and the aggregate structures/outputs shape used by the earlier CLI prototype. Multi-frame Extended XYZ input is also supported by mgl predict, mgl relax, and mgl md, with per-frame predictions, trajectory-preserving relaxation output, and one MD run per frame. The training and evaluation commands support Lightning accelerator/device selection, optional charge or magnetic-moment targets, dataset caching, and explicit stress units. Relaxation now exposes the ASE optimizer, cell-relaxation toggle, force threshold, and step limit. Model arguments accept local save paths, parser construction no longer queries the model registry, and MD boolean/mask arguments use unambiguous parsers.
• New: Release of compact ~1M parameter CHGNet MatPES models. Released lightweight (1,083,842 parameter) CHGNet foundation potentials for both PBE (materialyze/CHGNet-PES-MatPES-PBE-1M-2026.9) and r2SCAN (materialyze/CHGNet-PES-MatPES-r2SCAN-1M-2026.9) trained on the official MatPES 2025.2 dataset. These are now the default CHGNet models; the 2.7M materialyze/CHGNet-PES-MatPES-{PBE,r2SCAN}-2025.2.10 checkpoints remain available. Despite having ~2.5× fewer parameters than the standard 2.7M architecture, these compact models achieve strong train, validation, and test MAEs across energy (test: 26.72 meV/atom PBE, 27.45 meV/atom r2SCAN; val: 25.60 meV/atom PBE, 28.00 meV/atom r2SCAN; train: 22.58 meV/atom PBE, 25.46 meV/atom r2SCAN), forces (test: 110.53 meV/Å PBE, 137.58 meV/Å r2SCAN; val: 111.00 meV/Å PBE, 141.18 meV/Å r2SCAN; train: 86.75 meV/Å PBE, 112.27 meV/Å r2SCAN), and stresses (test: 0.6010 GPa PBE, 0.7094 GPa r2SCAN; val: 0.6060 GPa PBE, 0.7187 GPa r2SCAN; train: 0.4852 GPa PBE, 0.6343 GPa r2SCAN).
• Fix: M3GNet three-body messages were routed to the wrong bonds whenever threebody_cutoff < cutoff. create_line_graph / create_line_graph_torch enumerated triplets on the bond list pruned to threebody_cutoff, but ThreeBodyInteractions used those indices directly against parent-graph tensors, so each triplet took atom k and the cutoff weights f_c(r_ij) f_c(r_ik) from the wrong bonds and was scattered onto the wrong bond. This affected every MatGL release since v0.1.0 (DGL and PyG backends, and the LAMMPS export) with the default 5 Å / 4 Å cutoffs; models with threebody_cutoff == cutoff were unaffected. New M3GNet MatPES models have been refitted and released.
• Fix: CHGNet three-body geometry autograd detachment (#834). Continuous line-graph geometry features (lg_bond_vec and lg_bond_dist) were previously sliced under torch.no_grad(), causing three-body angular contributions to forces and stresses to be detached from autograd. Discrete graph topology is now isolated in torch.no_grad() while coordinate slicing preserves gradient tracking (@​wakamiya0315, @​bowen-bd).
• Fix: Line-graph periodic self-image supercell invariance (#839). Periodic self-image bonds meeting at a central atom in small unit cells were previously filtered and signed inconsistently in the line graph, causing a discrepancy between unit cell and supercell representations. Simplified edge connection logic and consistent bond vector inversion ensure exact supercell invariance across all cell dimensions. Updated the published 1M CHGNet PBE and r2SCAN models on Hugging Face Hub with the fix.
• Fix: Multi-GPU DDP training metric device mismatch and cache race condition. Fixed an issue where dummy metric tensors in PotentialLightningModule.loss_fn were constructed on CPU, causing NCCL sync_dist=True to crash, and guarded dataset cache directory cleanup in MGLDataset against multi-rank race conditions.
• LAMMPS pair_matgl fixes and speed-up (#825, #828, #831). The CMake snippets now register the matgl and matgl/kk pair styles, link libtorch directly for Kokkos, and no longer hardcode cluster-specific

... (truncated)

Changelog

Sourced from matgl's changelog.

4.1.0

• Disk-backed PyG training for datasets larger than memory. write_mgl_shards streams records into versioned, transactional CPU shards; MGLDiskDataset loads one shard per worker on demand; and ShardBatchSampler keeps batches shard-local while assigning disjoint shards and equal step counts to distributed ranks. MGLDataLoader now selects this path automatically for disk datasets without changing its existing in-memory behavior. MGLDataModule provides Lightning fit/validate/test/predict loaders with deterministic epoch shuffling, multiworker support, automatic MatGL collation, and optional on-the-fly graph conversion. Rebuild failures retain the prior committed manifest and clean up incomplete shards, while successful rebuilds remove superseded shards. A runnable QET notebook demonstrates sharded dataset creation, Lightning training, and per-graph QEq charge conservation.
• Improved PyG command-line workflows. mgl train now trains or fine-tunes interatomic potentials from local MatPES-shaped JSON/JSONL or Extended XYZ data through MGLDatasetLoader and MGLPotentialTrainer; mgl evaluate evaluates a saved potential with force/stress autograd enabled. Extended XYZ input supports periodic structures and nonperiodic molecules, standard or user-selected label keys, and TensorNet scratch training. The JSON loader accepts both current record-oriented MatPES files and the aggregate structures/outputs shape used by the earlier CLI prototype. Multi-frame Extended XYZ input is also supported by mgl predict, mgl relax, and mgl md, with per-frame predictions, trajectory-preserving relaxation output, and one MD run per frame. The training and evaluation commands support Lightning accelerator/device selection, optional charge or magnetic-moment targets, dataset caching, and explicit stress units. Relaxation now exposes the ASE optimizer, cell-relaxation toggle, force threshold, and step limit. Model arguments accept local save paths, parser construction no longer queries the model registry, and MD boolean/mask arguments use unambiguous parsers.
• New: Release of compact ~1M parameter CHGNet MatPES models. Released lightweight (1,083,842 parameter) CHGNet foundation potentials for both PBE (materialyze/CHGNet-PES-MatPES-PBE-1M-2026.9) and r2SCAN (materialyze/CHGNet-PES-MatPES-r2SCAN-1M-2026.9) trained on the official MatPES 2025.2 dataset. These are now the default CHGNet models; the 2.7M materialyze/CHGNet-PES-MatPES-{PBE,r2SCAN}-2025.2.10 checkpoints remain available. Despite having ~2.5× fewer parameters than the standard 2.7M architecture, these compact models achieve strong train, validation, and test MAEs across energy (test: 26.72 meV/atom PBE, 27.45 meV/atom r2SCAN; val: 25.60 meV/atom PBE, 28.00 meV/atom r2SCAN; train: 22.58 meV/atom PBE, 25.46 meV/atom r2SCAN), forces (test: 110.53 meV/Å PBE, 137.58 meV/Å r2SCAN; val: 111.00 meV/Å PBE, 141.18 meV/Å r2SCAN; train: 86.75 meV/Å PBE, 112.27 meV/Å r2SCAN), and stresses (test: 0.6010 GPa PBE, 0.7094 GPa r2SCAN; val: 0.6060 GPa PBE, 0.7187 GPa r2SCAN; train: 0.4852 GPa PBE, 0.6343 GPa r2SCAN).
• Fix: M3GNet three-body messages were routed to the wrong bonds whenever threebody_cutoff < cutoff. create_line_graph / create_line_graph_torch enumerated triplets on the bond list pruned to threebody_cutoff, but ThreeBodyInteractions used those indices directly against parent-graph tensors, so each triplet took atom k and the cutoff weights f_c(r_ij) f_c(r_ik) from the wrong bonds and was scattered onto the wrong bond. This affected every MatGL release since v0.1.0 (DGL and PyG backends, and the LAMMPS export) with the default 5 Å / 4 Å cutoffs; models with threebody_cutoff == cutoff were unaffected. New M3GNet MatPES models have been refitted and released.
• Fix: CHGNet three-body geometry autograd detachment (#834). Continuous line-graph geometry features (lg_bond_vec and lg_bond_dist) were previously sliced under torch.no_grad(), causing three-body angular contributions to forces and stresses to be detached from autograd. Discrete graph topology is now isolated in torch.no_grad() while coordinate slicing preserves gradient tracking (@​wakamiya0315, @​bowen-bd).
• Fix: Line-graph periodic self-image supercell invariance (#839). Periodic self-image bonds meeting at a central atom in small unit cells were previously filtered and signed inconsistently in the line graph, causing a discrepancy between unit cell and supercell representations. Simplified edge connection logic and consistent bond vector inversion ensure exact supercell invariance across all cell dimensions. Updated the published 1M CHGNet PBE and r2SCAN models on Hugging Face Hub with the fix.
• Fix: Multi-GPU DDP training metric device mismatch and cache race condition. Fixed an issue where dummy metric tensors in PotentialLightningModule.loss_fn were constructed on CPU, causing NCCL sync_dist=True to crash, and guarded dataset cache directory cleanup in MGLDataset against multi-rank race conditions.
• LAMMPS pair_matgl fixes and speed-up (#825, #828, #831). The CMake snippets now register the matgl and matgl/kk pair styles, link libtorch directly for Kokkos, and no longer hardcode cluster-specific

... (truncated)

Commits

• 8def32f Update uv.lock
• cfe632f Update version.
• 00ded2a docs: update 4.0.4 changelog and fix release notes extraction
• 013b63c Make 1M MatPES CHGNet models the default CHGNet (#844)
• 31372c6 fix(chgnet): fix directed line graph periodic self-image supercell in… (#843)
• 4483d56 Add disk-backed PyG datasets for large-scale training (#842)
• e4e8d70 feat(cli): Extending CLI functionalities including training and evaluation (#...
• d0b6bf0 Fix M3GNet three-body parent-bond index mapping when threebody_cutoff < cutof...
• 8816c92 Fix bugs in building the Kokkos implementation of LAMMPS (#831)
• 57336ad chore(deps): bump anyio from 4.14.1 to 4.14.2 (#837)
• 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)