OpenMMException error under MacOS

[PatWalters]

I made this change in line 1167 of helper_functions.py

from:
model_dict = torch.load(model_path)["model"]
to:
model_dict = torch.load(model_path,map_location=torch.device('cpu'))["model"]

However, when I try to run, I then get this OpenMMException.

---------------------------------------------------------------------------
OpenMMException                           Traceback (most recent call last)
[/var/folders/0n/pn_zp11569sfv4m2w6gzpwyr0000gp/T/ipykernel_13815/2831125247.py](http://localhost:8875/var/folders/0n/pn_zp11569sfv4m2w6gzpwyr0000gp/T/ipykernel_13815/2831125247.py) in ?()
      5 mol = Chem.MolFromSmiles('COCCO')
      6 mol = Chem.AddHs(mol)
      7 AllChem.EmbedMultipleConfs(mol, numConfs=128)
      8 
----> 9 minimized_mol, energies = minimize_mol(mol,"DMSO")
     10 entropies, free_energies = calculate_entropy(minimized_mol,"DMSO")

[~/software/GNNImplicitSolvent/Simulation/helper_functions.py](http://localhost:8875/~/software/GNNImplicitSolvent/Simulation/helper_functions.py) in ?(mol, solvent, model_path, solvent_dict, return_traj, tolerance, max_iterations, gnn_sim, return_gnn_sim, strides, cache, save_name, partial_charges, forcefield, constraints)
    989     num_iters = strides
    990     num_confs = num_rep // strides
    991 
    992     if gnn_sim is None:
--> 993         gnn_sim = get_gnn_sim(
    994             mol=mol,
    995             solvent=solvent,
    996             model_path=model_path,

[~/software/GNNImplicitSolvent/Simulation/helper_functions.py](http://localhost:8875/~/software/GNNImplicitSolvent/Simulation/helper_functions.py) in ?(mol, solvent, model_path, solvent_dict, cache, save_name, partial_charges, forcefield, constraints, num_confs)
   1164             constraints=constraints,
   1165         )
   1166     else:
   1167         model_dict = torch.load(model_path,map_location=torch.device('cpu'))["model"]
-> 1168         gnn_sim = create_gnn_sim(
   1169             smiles,
   1170             cache=cache,
   1171             num_confs=num_confs,

[~/software/GNNImplicitSolvent/Simulation/helper_functions.py](http://localhost:8875/~/software/GNNImplicitSolvent/Simulation/helper_functions.py) in ?(smiles, cache, num_confs, workdir, run_name, save_name, rdkit_mol, solvent_dict, solvent, num_solvents, model_dict, solvent_model, solvent_dielectric, partial_charges, forcefield, constraints)
    818         constraints=constraints,
    819     )
    820 
    821     gnn_sim._ref_system.platform = "GPU"
--> 822     gnn_sim.setup_replicates(only_check_first=True)
    823     return gnn_sim

[~/software/GNNImplicitSolvent/Simulation/Simulator.py](http://localhost:8875/~/software/GNNImplicitSolvent/Simulation/Simulator.py) in ?(self, only_check_first)
   1590         # test whether forces are equal
   1591         self._ref_system.create_system()
   1592         self._ref_system.create_simulation(self._datahandler.positions)
   1593         self._ref_system.platform = "GPU"
-> 1594         ref_single_forces = self._ref_system.calculate_forces()
   1595 
   1596         # Create copies of the system
   1597         m = Modeller(self._datahandler.topology, self._datahandler.positions)

[~/software/GNNImplicitSolvent/Simulation/Simulator.py](http://localhost:8875/~/software/GNNImplicitSolvent/Simulation/Simulator.py) in ?(self, adapt_values, use_NN)
    788             radii_and_charge_data[:, 1] = scale
    789             if adapt_values:
    790                 self.adapt_values(radii_and_charge_data)
    791 
--> 792         forces = self._simulation.context.getState(getForces=True).getForces()
    793         x_forces = np.array([force[0] for force in forces._value])
    794         y_forces = np.array([force[1] for force in forces._value])
    795         z_forces = np.array([force[2] for force in forces._value])

[~/miniforge3/envs/GNNImplicitSolvent_run/lib/python3.10/site-packages/openmm/openmm.py](http://localhost:8875/~/miniforge3/envs/GNNImplicitSolvent_run/lib/python3.10/site-packages/openmm/openmm.py) in ?(self, getPositions, getVelocities, getForces, getEnergy, getParameters, getParameterDerivatives, getIntegratorParameters, enforcePeriodicBox, groups)
   2407         if getParameterDerivatives:
   2408             types += State.ParameterDerivatives
   2409         if getIntegratorParameters:
   2410             types += State.IntegratorParameters
-> 2411         state = _openmm.Context_getState(self, types, enforcePeriodicBox, groups_mask)
   2412         return state

OpenMMException: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript, serialized code (most recent call last):
  File "code[/__torch__/MachineLearning/GNN_Models/___torch_mangle_10.py", line 391](http://localhost:8875/__torch__/MachineLearning/GNN_Models/___torch_mangle_10.py#line=390), in forward
      size_i0 = _115[0]
    x = torch.cat([x_i0, x_j0, edge_attributes1], 1)
    _145 = torch.add(torch.matmul(x, CONSTANTS.c14), CONSTANTS.c15)
                     ~~~~~~~~~~~~ <--- HERE
    x0 = torch.silu(_145)
    _146 = torch.add(torch.matmul(x0, CONSTANTS.c16), CONSTANTS.c17)

Traceback of TorchScript, original code (most recent call last):
RuntimeError: expected m1 and m2 to have the same dtype, but got: double != float

[katzberger]

@PatWalters Thank you very much for your interest in our work!

I apologize for the difficulty in getting this up and running. Unfortunately, I am on vacation, which makes my responses slower. In our research, we have only run this code on GPUs, so we did not encounter these issues.

Nevertheless, running it on CPUs should pose no problem (even though the speed might be reduced).

I have quickly looked into the error, and casting the atomic positions to float seems to resolve it. I have created a new branch, CPU_MacOS, with the necessary adjustments, and this now seems to work on MacOS. Please let us know if that resolved the problem for you as well

In any case thank you for your patience.

[PatWalters]

Thanks, I'll clone that branch and give it a try.