After this is merged, we should cut 0.17.0!

Quick timing update for c-Met with 133 states (29 different electrostatic discharging states):

Old code:

Computing energy matrix took    2.485s

New code:

Computing energy matrix took    0.413s

Quite impressive!

Do we have to add the new classes (such as state.GlobalParameterFunction) to docs/states.rst for the API docs to be rendered?

I think you're right. I'll add them before merging.

Does the test_overlap() test pass locally?

I'll run all the alchemy.py tests on the cluster to make sure.

After this is merged, we should cut 0.17.0!

I'd wait for OpenMM 7.3 release before releasing this feature and test the dev openmmtools in the meantime. I'm back on debugging the NaNs today.

Also, I still would like to consider keeping the order of lambda consistent between different models with the addition of a sqrt_lambda_electrostatics parameter. I think the changes to the code would be minimal, just changing the name of the offset parameter and an addition to AlchemicalState like

class AlchemicalState:
    ...
    _sqrt_lambda_electrostatics = GlobalParameter('sqrt_lambda_electrostatics', expression='sqrt(lambda_electrostatics)')

I'd work on it on a separate PR in any case.

@andrrizzi : Can this be merged and a new release cut?

I still have to run the full test_alchemy.py suite before merging. Also, I'd wait for NaN issue on OpenMM 7.3 to be fixed before making a release since we'll have to pin the minimum openmm version. That's my first priority coding-wise.

Thanks for clarifying!

I've run all the tests in test_alchemy.py and they passed. AppVeyor is failing because of problems with the win build of openmm7.3rc, but I think we can merge this for now.

Hooray!

@andrrizzi : Maybe we should change the YANK default for alchemical_pme_treatment: to exact now?