Error when running simulation using acemd

[AminaMnhr]

Hi,

I am trying to use acemd to run simulations but I get into two errors when using acemd to run simple MD or NN/MM simulations.

Here is how I installed acemd:
module load python/3.10.13
module load cuda/12.3.1
conda install htmd acemd cuda-version=12.3 -c acellera -c conda-forge

Problem number 1 : running MD simulation with acemd:

I am trying to use acemd to run a MD simulation as explained in this tutorial: https://software.acellera.com/acemd/tutorial.html
using the API Python.
However when I run the command:
acemd("/tmp/my_simulation"), it works and the simulation works properly.
But when I run specifying the run variable:
acemd("/tmp/my_simulation", run="2ns")
It does not work, and I get a problem of 'str variable' when specifying the run.

Problem number 2 : running NN/MM simulation with acemd

I tried to run the NN/MM example here : https://software.acellera.com/acemd/nnpmm.html
I did my hugging account and got the NN Aceforce on a cluster.

But when I run the acemd command to launch the NN/MM simulation, I get the error when reading the NN:

AssertionError: Torch not compiled with CUDA enabled

Problem number 3 : NN pipeline with acemd

• It is my understanding that I could either launch my simulations using htmd SlurmQueue or directly the acemd command, if I want to compute RBFE using the NN/MM. What would you recommand, using htmd with a SlurmQueue with s.submit or directly using acemd command ?

• For using the NN, we specify the name, file and type in the input.yaml file. But is that for the input file of the equilibration part or the production part of the simulation? Or both?

• Lastly, when I use htmd to set up the equilibration part or production part (as in the tutorial), the setup_equilibration command and setup_production creates automatically for me the input.yaml. So if I want to use a NN, do I have to modify this input.yaml after the setup command, adding the nn part directly?

Thanks a lot for your help!

[stefdoerr]

Yes about problem number 1, I am aware of the issue. You can only override the runtime with an integer number of steps currently. I might fix it in future versions.

Problem number 2: your conda env which is executing the job on the cluster doesn't have pytorch with CUDA support compiled. I am assuming that maybe you are using a different environment from the one you installed acemd in. You can check it by running conda list pytorch in the cluster and see the version it prints, it will probably include cpu in the string instead of cuda.

Problem number 3:

• You can use directly a SLURM command if you are familiar with SLURM. I have not tested yet ACEMD4 with SlurmQueue although I assume it should work
• Yes, if you want to do both equil and prod with NNP/MM then include the nnp section in both input files
• Yes you need to modify the automatically created input.yaml files and add the nnp section to them

[AminaMnhr]

Thanks a lot for your fast and clear answer!

Problem number 2:

I activated my acemd environment which contain cuda 12.3, as well as pytorch :
(acemd) amenhour@balam-login01:~/Example_NNP_MM_input_file$ conda list pytorch

Name Version Build Channel

pytorch 2.5.1 cpu_mkl_py312_heeca0f5_108 conda-forge
pytorch-lightning 2.5.0.post0 pyh101cb37_0 conda-forge
pytorch_geometric 2.6.1 pyhd8ed1ab_1 conda-forge

As you said, it print cpu. Also if I do import torch and torch.cuda.is_available() in my acemd environment, it returns False. Do you know how I fix this?
I did try to remove pytorch : conda remove torch torchaudio torchvision
And then: conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=12.1 -c pytorch -c nvidia
But same problem...

Problem number 3

Ok so If I use acemd to run my simulations and not htmd directly, is the pipeline still (for RBFE) :

• setup_equilibration command
• adding NN to input.yaml file
• running acemd command
  And then after the equilibration is done:
• setup_production command
• adding NN to input file
• running acemd command

It is my understanding that acemd command should then be runned twice (one for equilibration, and another for production) ? The minimization does not have to be runned like this?
Also is the equilibration using one GPU, and then the production will be using as well one GPU?

(Just for context: I am trying to compute RBFE from the QuantumbindRFPE paper, so I wonder if the pipeline has to be somewhat like this, splitting equil and production with two different input files.)

Thanks again!!

[stefdoerr]

On why it installed the CPU version I cannot know for sure. Normally if you run:
conda install htmd acemd cuda-version=12.3 -c acellera -c conda-forge
It should install the CUDA versions of the packages. But if you want you might be able to force it by running
conda install htmd acemd cuda-version=12.3 pytorch-gpu -c acellera -c conda-forge

Problem 3:
Yes this is the correct pipeline if you want to run ACEMD simulations. Minimization is included in the first part of the equilibration config so it's run then. However if you are trying to reproduce QB-RBFE, we didn't use ACEMD there, we used AToM-OpenMM which uses OpenMM directly. You won't be able to use the ATM force directly in ACEMD to reproduce QB-RBFE