GPU test_alchemy.py failures.
See original GitHub issueHello,
I am experiencing some test failures for OpenMMTools when using the GPU platform and was hoping I could get some help resolving them.
I have set up a test environment for OpenMMTools using a Conda environment and the YAML here: https://github.com/choderalab/openmmtools/blob/master/devtools/conda-envs/test_env.yaml which yields this environment for me: environment.txt
I git clone OpenMMTools, install with pip, then run nosetests test_alchemy.py
in the test directory of my clone.
On CPU this yield the following output: cpu.txt
On GPU I get many more errors about large energy differences for HostGuestExplicit with PME based tests: gpu.txt
Am I running these test correctly or am I missing something in the GPU case? Any help with this would be greatly appreciated.
For reference I’m using an Ubuntu Linux 64 machine with an RTX 2070S, Driver Version: 470.63.01, CUDA Version: 11.4, cudatoolkit=11.2.2
Thanks, Alex
Issue Analytics
- State:
- Created 2 years ago
- Comments:14 (14 by maintainers)
Hello,
I have run these tests again on my machine this morning using the latest OpenMM and OpenMMtools installed by conda which gives this env: environment.txt
I now no longer see any of these errors on GPU. For reference I’m now using CUDA Driver Version: 510.47.03 and CUDA Version: 11.6 with RTX2070S. So I think with the latest code this is no longer an issue and I will close it now, Thanks 😃
Alex
For one of the failing tests (ideally the simplest/smallest one), can you serialize out the
System
andState
objects to XML for us to investigate? Something like this:This will make it easier for us to look at in detail, especially if we have to bring in Peter Eastman.