feat: add double precision (float64) inference support#156
Merged
Conversation
yanghan234
force-pushed
the
feat/double-precision-inference
branch
2 times, most recently
from
52198e1 to
8aa2863
Compare
yanghan234
added a commit
that referenced
this pull request
May 11, 2026
float64 support belongs to PR #156. For now use the wrapper's default float32 in model_loading and doc examples. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
yanghan234
added a commit
that referenced
this pull request
May 11, 2026
* feat: add load_mattersim helper to forcefield Add a convenience function that wraps Potential.from_checkpoint() with gradient checkpointing support and automatic version normalization from checkpoint paths. Exported from mattersim.forcefield. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * feat: add torch-sim-atomistic as required dependency Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * feat: add torchsim integration for batch MD and relaxation Port TorchSim support from the internal repo, providing GPU-accelerated batch molecular dynamics and structure relaxation via torch-sim. New package: mattersim.torchsim - TorchSimWrapper: adapts MatterSim Potential as a TorchSim ModelInterface - TorchSimBatchMD: batch MD with NVT/NPT integrators, temperature schedules - TorchSimBatchRelaxer: batch structure optimization with FIRE/etc. - Settings classes (IntegratorSettings, OptimizerSettings) with validation - Trajectory loading/saving via TorchSim H5MD format - Graph construction bridge (build_graph_from_simstate) - Model loading with AOTI compilation support Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * test: add torchsim integration tests 27 tests covering settings validation, TorchSimWrapper forward pass, batch relaxation (file + in-memory), batch MD with multiple integrators, trajectory continuation, per-system temperature schedules, and trajectory loader convenience constructors. GPU-heavy tests are marked @slow and @requires_gpu so they are skipped on CPU-only machines. Also adds mattersim_potential_best_device fixture to conftest.py. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix: set version in load_mattersim for autobatcher lookup load_mattersim now always sets Potential.version so the pre-computed autobatcher memory scaler table is used instead of falling back to slow runtime memory estimation. - Default checkpoint (load_path=None) gets 'mattersim-v1.0.0-1M' - Explicit paths are normalized to canonical version strings Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * chore: clean up dead python_version < 3.10 conditionals Since requires-python is now >=3.12 (for torch-sim-atomistic), the python_version < 3.10 dependency conditionals for emmet-core and numpy are unreachable. Simplify to unconditional versions. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * ci: update test matrix to Python 3.12+ only Drop Python 3.10 from CI matrix since requires-python is now >=3.12. Add Python 3.13 to the test matrix. Update macOS job to 3.12. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * ci: drop Python 3.10/3.11 wheel builds from release pipeline Only build wheels for Python 3.12 and 3.13 since torch-sim-atomistic requires Python >=3.12. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * ci: bump gh-pages docs build to Python 3.12 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * docs: update Python version requirement to 3.12+ in README Update badge, prerequisite, and conda create example to reflect the new minimum Python 3.12 requirement. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * chore: update Python version to 3.12+ in MODEL_CARD, environment.yaml, cibuildwheel Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * refactor: remove batch MD and relaxation runners from torchsim Remove the batch runner infrastructure (TorchSimBatchRunner, TorchSimBatchMD, TorchSimBatchRelaxer), their settings classes (IntegratorSettings, OptimizerSettings), trajectory loaders, and associated tests. Keep the core TorchSim wrapper (TorchSimWrapper, graph_construction, model_loading) that adapts MatterSim potentials as TorchSim ModelInterface. Deleted files: - src/mattersim/torchsim/base.py - src/mattersim/torchsim/batch_relax.py - src/mattersim/torchsim/md.py - src/mattersim/torchsim/settings.py - src/mattersim/torchsim/settings_base.py - src/mattersim/torchsim/trajectory_loader.py Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * docs: add TorchSim simulation guide Add user guide for running structure relaxation and molecular dynamics using the TorchSim backend with MatterSim potentials. Covers wrapper creation, optimization, MD with various integrators, and trajectory saving. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * docs: fix convergence function API in TorchSim guide Use correct parameter name 'force_tol' (not 'fmax') and drop 'include_cell_forces' from the minimal example. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix: remove unused emmet-core dependency emmet-core is not imported anywhere in the codebase. Removing the dead dependency per review feedback. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * bump torchsim >= 0.6.0 * fix: remove remaining load_mattersim references Replace load_mattersim() calls in model_loading.py with direct Potential.from_checkpoint() + enable_gradient_checkpointing() calls. Remove load_mattersim from forcefield __init__.py exports. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix: remove _sanitize_outputs and sanitize_nan plumbing Remove internal-only NaN sanitization logic per review feedback. Removed from TorchSimWrapper and get_torchsim_wrapper. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix: drop explicit float64, use default float32 dtype float64 support belongs to PR #156. For now use the wrapper's default float32 in model_loading and doc examples. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
yanghan234
force-pushed
the
feat/double-precision-inference
branch
from
8aa2863 to
8a56b46
Compare
Add dtype='float64' option to MatterSimCalculator for double-precision
inference. This improves numerical stability for sensitive properties
like thermal conductivity.
Changes:
- MatterSimCalculator: new dtype parameter ('float32' or 'float64'),
converts model weights via model.double() and creates inputs in the
matching dtype
- m3gnet.py: replace .float() with .to(pos.dtype) for pbc_offsets so
it respects the model's precision
- calculate(): upcast legacy-path float tensors to match model dtype
On equilibrium Si diamond, float64 reduces force noise from ~1e-6 to
~1e-15 eV/Å compared to float32.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Switch BTE tests to dtype='float64' for more stable force constants. Update reference kappa values to float64 means verified across 3 runs (rel_std < 0.03%). Tighten tolerance from 8% to 2%. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
MPS does not support float64 tensors. This commit: - Validates dtype+device combo early in MatterSimCalculator.__init__, raising ValueError before loading the model - Falls back to float32 in the basic BTE test on MPS - Skips the strict BTE test on MPS (reference values require float64) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Use float32 with 10% kappa tolerance and 0.5% FC tolerance on MPS, while keeping float64 with tight tolerances (2% / 0.1%) on CUDA. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
yanghan234
force-pushed
the
feat/double-precision-inference
branch
from
8a56b46 to
e145b41
Compare
The local 'import torch._inductor.config' statement caused Python to treat 'torch' as a local variable in __init__, making earlier references to torch.float32/torch.float64 fail with UnboundLocalError. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Collaborator
Author
|
@copilot review this PR |
Agent-Logs-Url: https://github.com/microsoft/mattersim/sessions/b2ac740c-f04e-4e6b-a5c0-b3389a4caaf3 Co-authored-by: yanghan234 <32856137+yanghan234@users.noreply.github.com>
Contributor
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds a
dtypeparameter ("float32"or"float64") toMatterSimCalculatorfor double-precision inference. This significantly improves numerical stability for sensitive properties like thermal conductivity.On equilibrium Si diamond, float64 reduces force noise from ~1e-6 to ~1e-15 eV/Å compared to float32.
Changes
src/mattersim/forcefield/potential.pydtypeparameter onMatterSimCalculator.__init__with validationmodel.double()whendtype="float64"is selectedsrc/mattersim/forcefield/m3gnet/m3gnet.pypbc_offsets.float()→pbc_offsets.to(pos.dtype)to respect the model's precisiontests/applications/test_bte.pydtype="float64"for reproducibility