Iván Pulido

I see, yeah I had the feeling it was a storage issue. We probably need to clean things up in general but I think this will help. Thanks for clarifying...

@mikemhenry Yes, I agree we probably need to discuss first what we want to store and when. And which of these are needed when resuming simulations and similar situations. If...

> the time series data we save for analysis should be user configurable. Yes, decoupling the _checkpointing_ with the user-configurable time series data would probably work better here. I'd vote...

I wonder if this is what's biting us here, though I don't know why are we expecting these arrays to be masked? https://github.com/choderalab/openmmtools/pull/701

Now, to be fair, I don't think the changes in that linked PR have anything to do with the serialized/deserialized positions or velocities. So I doubt it.

I'll be also running a pass of the perses tyk2 benchmark just to check this doesn't change the results from it.

I found out that with these changes the minimization seems to be significantly slower and is not using the 100% (nor close) of the GPU. For the `tyk2` benchmark, the...

In case it helps, the `LocalEnergyMinimizer` ends up using the `GeodesicBAOABIntegrator` from OpenMM in the end.

FYI, I'm having a bit of trouble understanding the `tolerance` units here, we are passing `Quantity(value=1.0, unit=kilojoule/(nanometer*mole))` as tolerance, but as far as I understand, tolerance for energy should just...

Timings are as follows (initial and final energies are just for the first replica): | Minimizer | tolerance | time minimizing all reps. | initial E kT | Final E....