OpenMM-Torch 1.5

This release is for use with OpenMM 8.2. Major new features include support for calculating derivatives of the energy with respect to global parameters, and support for the HIP platform. In addition, the OpenCL platform will now use the GPU to compute the PyTorch model if CUDA or HIP is available.

What's Changed

• Can compute parameter derivatives by @peastman in #143
• Attempts at fixing CI by @peastman in #144
• Freeze pytorch model by @peastman in #148
• Add energy derivatives to README by @RaulPPelaez in #145
• Serialize property values by @peastman in #152
• Update Mac build to newer versions by @peastman in #153
• Created HIP platform by @peastman in #157
• Add pytorch lib dir to rpath by @peastman in #158
• Don't assume SSE is present by @peastman in #160
• Changes for PyPI wheels by @peastman in #161

Full Changelog: v1.4...v1.5

OpenMM-Torch 1.4

This release fixes an issue with CUDA Graphs (#122) and support for Python 3.12 (#123).

This release can be installed from conda-forge:

conda install -c conda-forge openmm-torch=1.4

What's Changed

• Use the same stream as the OpenMM context when replaying a CUDA graph. by @RaulPPelaez in #122
• Use setuptools instead of distutils by @peastman in #123

Full Changelog: v1.3...v1.4

OpenMM-Torch 1.3

This release fixes an issue with CUDA Graphs for the backward pass (#120).

This release can be installed from conda-forge:

conda install -c conda-forge openmm-torch=1.3

What's Changed

• Add AWS GPU Runner by @mikemhenry in #107
• Fix CUDA Graphs for the backward pass by @RaulPPelaez in #120

New Contributors

• @mikemhenry made their first contribution in #107

Full Changelog: v1.2...v1.3

OpenMM-Torch 1.2

This release fixes (#116) the issue of failing minimization with NNP due to a device change (openmm/NNPOps#112).

This release can be installed from conda-forge:

conda install -c conda-forge openmm-torch=1.2

What's Changed

• Document the constructor that takes a scripted module by @RaulPPelaez in #104
• Clone the module on TorchForceImpl initialization by @RaulPPelaez in #116

Full Changelog: v1.1...v1.2

OpenMM-Torch 1.1

This release brings support for PyTorch 2 (#106) and CUDA Graphs (#103).

This release can be installed from conda-forge:

conda install -c conda-forge openmm-torch=1.1

What's Changed

• Attempt at fixing SWIG issues by @peastman in #96
• Add a constructor to TorchForce that takes a torch::jit::Module by @RaulPPelaez in #97
• Use Ubuntu-22 instead of deprecated Ubuntu 18 in CI by @RaulPPelaez in #105
• Making TorchForce CUDA-graph aware by @RaulPPelaez in #103
• Torch2 compatibility by @RaulPPelaez in #106

New Contributors

• @RaulPPelaez made their first contribution in #97

Full Changelog: v1.0...v1.1

OpenMM-Torch 1.0

This release (1.0) marks that OpenMM-Torch has matured to be used for production. We aim to maintain a stable and backward-compatible API for the 1.x releases.

This release can be installed from conda-forge:

conda install -c conda-forge openmm-torch=1.0

What's Changed

• Use Mamba for CI by @raimis in #81
• Update CI for PyTorch 1.12 by @raimis in #83
• Pin swig to <4.1 by @raimis in #86

Full Changelog: v0.8...v1.0

OpenMM-Torch 0.8

This release fixes a bug (#80) which prevented TorchForce from being used within OpenMM.CustomCVForce.

This release can be installed from conda-forge:

conda install -c conda-forge openmm-torch

What's Changed

• Fix the tutorial by @raimis in #77
• Push primary context before invoking PyTorch by @peastman in #78
• Revert "Push primary context before invoking PyTorch" by @raimis in #79
• Fix interoperability with CustomCVForce by @raimis in #80

Full Changelog: v0.7...v0.8

OpenMM-Torch 0.7

This release fixes two packaging issues. No new features have been implemented.

This version can be installed from conda-forge via

conda install -c conda-forge openmm-torch

What's Changed

• Add PyTorch lib directory to library path by @peastman in #72
• Update the test models and CI dependencies by @raimis in #73

Full Changelog: v0.6...v0.7

OpenMM-Torch 0.6

This release fixes an issue where the PyTorch module was always loaded to the GPU device 0, regardless which device OpenMM was set to use (#70). Also, the Python API has been fixed by including TorchForce.getOutputsForces and TorchForce.setOutputsForces (#60).

A new tutorial illustrating Openmm-Torch with the accelerated kernel library NNPOps is available:

This version can be installed from conda-forge via

conda install -c conda-forge openmm-torch

What's Changed

• Fix the Python wrapper of "setOutputsForces" and "getOutputsForces" by @raimis in #60
• Move the PyTorch module to a correct device by @raimis in #70
• The first tutorial of OpenMM-Torch with NNPOps by @raimis in #62

Full Changelog: v0.5...v0.6

OpenMM-Torch 0.5

This release adds the ability to output the computed atomic force directly (#52) without using the back-propagation of energy; and fixes a long-standing issue the CUDA context synchronization (#47, #49), which was causing crashes or incorrect results. All the users are advised to update.

What's Changed

• Improved coordination of CUDA contexts with PyTorch by @peastman in #47
• Update CI by @raimis in #53
• Test if a PyTorch module receives corrects arguments by @raimis in #50
• Models can directly output forces by @peastman in #52
• Fix the synchronisation between the CUDA contexts by @raimis in #49

Full Changelog: v0.4...v0.5