GPU test_alchemy.py failures.

Hello,

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

[adw62]

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

[jchodera]

For one of the failing tests (ideally the simplest/smallest one), can you serialize out the System and State objects to XML for us to investigate? Something like this:

from simtk.openmm import XmlSerializer
with open('system.xml', 'wt') as outfile:
    outfile.write(XmlSerializer.serialize(context.getSystem()))
with open('state.xml', 'wt') as outfile:
    outfile.write(XmlSerializer.serialize(context.getState(getPositions=True, getForces=True, getEnergy=True, getParameters=True, getIntegratorParameters=True)))

This will make it easier for us to look at in detail, especially if we have to bring in Peter Eastman.