* Try to fix CI on Linux
* Don't install POCL on Mac
* Don't install khronos-opencl-icd-loader
* Update OpenMM version
* Updated tested versions
* Fixed version number
* Install correct packages for CUDA 12
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Try different method of installing CUDA
* Try not installing CUDA packages from conda
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Don't build OpenCL on Mac
* Don't try to run tests that can't run correctly
* Update C++ version and minimum macOS version
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Debugging
* Fixed working directory for tests
* Fixes to testing
* Try using PyTorch 2.3
* Try using PyTorch 2.2
* Switch back to PyTorch 2.1
* Add CUDA graph draft
* Initialize energy and force tensors in the GPU.
* Add comment on graph capture
* Catch torch exception if the model fails to capture.
* Replay graph just after construction
Finish capturing before rethrowing if an exception occurred during capture
* Add python-side test script for CUDA graphs
* Implement properties
* Update the Python bindings
* Unify the API for properties
* Pass the propery map to the constructor
* Skip graph tests if no GPU is present
* Guard CUDA graph behavior with the CUDA_GRAPH_ENABLE macro
* Check validity of the useCUDAGraphs property
* Add missing bracket to openmmtorch.i
* Fix bug in useCUDAgraph selection
* Update tests
* Add test for get/setProperty
* Update documentation with new functionality
* Add a CUDA graph test for a model that returns only energy
* Add contributors
* Reset pos grads after graph capture. Make energy and force tensors persistent.
* Add tests that execute the model many times to catch bugs related with
CUDA graph capture
* Run formatter
* Warmup model for several steps
* Include gradient reset into the graph
* Do not reset energy and force tensors before graph capture
* Remove unnecessary line
* Add tests for larger number of particles
* Remove unnecessary compilation guard now that Pytorch 1.10 is not supported
* Simplify getTensorPointer now that Pytorch 1.7 is not supported
* Change addForcesToOpenMM to addForces
* Change execute_graph to executeGraph
* Wrap graph warming up in a try/catch block
* Add correctness test for modules that only provide energy
* Revert "Add correctness test for modules that only provide energy"
This reverts commit d20f4bfa83.
* Explicit conversion to correct type in getTensorPointer
* Added a new property for TorchForce, CUDAGraphWarmupSteps.
* Clarify docs
* Document properties
* Throw if requested property does not exist
* Change getProperty(string) to getProperties()
* Add getProperties to python wrappers
* Fix formatting
* Set default properties
* Update tests
* Update some comments
---------
Co-authored-by: Raimondas Galvelis <r.galvelis@acellera.com>
* Add version number as a member to TorchForceProxy
* Encode the model file contents when serializing TorchForce
* Add tests for new TorchForce serialization
* Fix test not finding Python executable
* Format include directives correctly
* Hardcode TorchForceProxy version number
* Fix formatting issues
* Move Python serialization test to the correct place
* Make function encodeFromFileName static
* Update serialization python test to correctly remove temporary files after executing
* Use the base64 encoding capabilities of openssl to serialize model file
* Update TorchForce serializer
* Add a constructor to TorchForce that takes a torch::jit::Module.
TorchForce(string fileName) is implemented by delegating to the new
constructor.
Update serialization test accordingly to compare the module file name
and the module itself.
* Remove unnecessary include
* Change i_file to file in TorchForce constructor
* Add swig typemaps to new TorchForce constructor
* Add setup.py as a dependency for the PythonInstall CMake rule
* Fix swig out typemap for torch::jit::Module
Now it is possible to call getModule() on a TorchForce object from
Python, which will return a module of the same type as, for instance, torch.jit.load()
* Remove commented line in CMakeLists.txt
* Remove unnecessary dependency in setup.py
* Add more tests for new constructor
* Add some comments for the new constructor
* Updates to TorchForce serialization
* Use hex encoding instead of base64 for serialization.
SSL no longer a direct dependency.
* Remove unnecessary header
* Update Python serialization test
* Minor changes
* Improve temporary path handling in python serialization tests
* More informative exception when failing to serialize TorchForce
* Remove unnecessary check in TorchForce serialization
* Changes to C++ serialization tests
* Changes to C++ serialization tests
* Add a test with CustomCVForce
* Test all the platforms
* Add an iteroperability test for TorchANI and NNPOps
* Add a missing dependencies
* Skip for MacOS
* Move imports
* Fix import
* Retain the primary context
* Switch properly the contexts
* Set the oldest CUDA to 11.0
* Fix nvcc version
* Enable an extra check
* Clean up a temporary file
* Add more checks
* Add comments
* Remove a sync and clean up
* Move the primary context activation
