issue while running pretrained mace model with openmm MACELES checkpoint runs in native MACE GPU but fails in OpenMM-ML CUDA with CPU/CUDA tensor mismatch in torch.functional.tensordot

[jagriti-iisc]

File "/.../torch/functional.py", line 1333, in tensordot
return _VF.tensordot(a, b, dims_a, dims_b)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

tensordot = torch.functional.tensordot(einsum_11, mul_15, ([3, 4], [0, 1]), out = None)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument mat2 in method wrapper_CUDA_mm)

Summary

A local MACELES checkpoint loads successfully and also runs correctly with native MACECalculator(device="cuda"), but fails during GPU execution inside OpenMM-ML with a CPU/CUDA tensor mismatch.

The model can be loaded with torch.load() and resolves to:

<class 'mace.modules.extensions.MACELES'>
``

[peastman]

Can you provide a self-contained example that reproduces the problem? Provide your complete script and all necessary input files. Also, can you post the output of conda list (if you installed with conda) or pip list (if you installed with pip)?

[jagriti-iisc]

Thanks for looking into this.

Below is a minimal, self-contained example reproducing the issue, along with
the environment details.

---

Minimal reproduction script (OpenMM‑ML + MACELES)

System:

• Ion-watersystem (~1000 atoms)
• Periodic box
• Single-point energy or short minimization is sufficient to trigger the error

Script:

from openmm.app import PDBFile, Simulation
from openmm.unit import *
import openmm as mm
from openmmml import MLPotential

pdb = PDBFile("mg_water_box.pdb")

potential = MLPotential(
    "mace",
    modelPath="Mg_Water_MACELES.model"
)

system = potential.createSystem(
    pdb.topology,
    precision="single",
    device="cuda"
)

integrator = mm.LangevinMiddleIntegrator(
    300*kelvin,
    1.0/picosecond,
    1.0*femtosecond
)

platform = mm.Platform.getPlatformByName("CUDA")
simulation = Simulation(pdb.topology, system, integrator, platform)
simulation.context.setPositions(pdb.positions)

# Any of the following triggers the error:
simulation.context.getState(getEnergy=True)
# or:
# simulation.minimizeEnergy()
``


Observed error
When running on CUDA, OpenMM‑ML fails with a PyTorch error similar to:

RuntimeError: Expected all tensors to be on the same device, but found
at least two devices, cuda:0 and cpu

Expected behavior
The same MACELES checkpoint:

runs correctly in native MACE (GPU)
runs correctly in native MACE (CPU)
runs correctly in OpenMM‑ML when forced to CPU (device="cpu")
Only OpenMM‑ML + CUDA triggers the tensor device mismatch.

Model details
Model type: MACELES (trained via current MACE training workflow)
Checkpoint format: standard .model saved by MACE
No custom modifications to the model code
Environment information
Installed via conda‑forge + pip split:

(consolidated output of conda list / pip list below)

Key versions:

Python: 3.10
OpenMM: 8.4.x
openmm‑ml: 1.6
PyTorch: 2.x (CUDA enabled)
CUDA runtime: 12.x
MACE: installed via mace‑torch

Thanks again for your time.

[peastman]

OpenMM: 8.4.x
openmm‑ml: 1.6

OpenMM-ML 1.6 requires OpenMM 8.5. How did you manage to install that combination? It's supposed to enforce it as a requirement. Can you post the complete output of conda list?

Try upgrading OpenMM.

[jagriti-iisc]

Date2026-04-24(i did the instalaltion again)
SystemUbuntu 22.04.5 LTS (kernel 6.8.0-88-generic)
CPUAMD Ryzen 9 7950X 16-Core (32 threads)
RAM124 GB
GPUNVIDIA GeForce RTX 4070 Ti SUPER (16 GB VRAM)
CUDA Driver570.195.03
CUDA Toolkit12.8
Conda Environmentmace_oom_claude
Python3.11 (conda-forge) (this is my system)

[peastman]

It says you're using OpenMM 8.2 and OpenMM-ML 1.2? That's different from what you said before, and they're very old versions.

Create a new environment:

conda create --name openmm python
conda activate openmm

Install the relevant packages into it:

pip install openmm[cuda12] openmm-ml mace-torch

And then see if it works. If it doesn't, post the complete output of conda list. Not just the packages you think are relevant, and not a screenshot that cuts off some columns.

[jagriti-iisc]

Thank you for the suggestion. pip install openmm-ml fails with "No matching distribution found" — it appears openmm-ml is not available on PyPI. Should I install openmm via pip and openmm-ml via conda-forge separately, or is there an alternative approach you'd recommend?

(base) jag@jag:$ conda activate omm-pip
(omm-pip) jag@jag:$ python -m pip install "openmm[cuda12]" openmm-ml mace-torch
ERROR: Could not find a version that satisfies the requirement openmm-ml (from versions: none)
ERROR: No matching distribution found for openmm-ml
(omm-pip) jag@jag:~$ conda list

packages in environment at /home/jag/anaconda3/envs/omm-pip:

Name Version Build Channel

_libgcc_mutex 0.1 main
_openmp_mutex 5.1 1_gnu
bzip2 1.0.8 h5eee18b_6
ca-certificates 2026.3.19 h06a4308_0
ld_impl_linux-64 2.44 h9e0c5a2_3
libexpat 2.7.5 h7354ed3_0
libffi 3.4.4 h6a678d5_1
libgcc 15.2.0 h69a1729_7
libgcc-ng 15.2.0 h166f726_7
libgomp 15.2.0 h4751f2c_7
libnsl 2.0.0 h5eee18b_0
libstdcxx 15.2.0 h39759b7_7
libstdcxx-ng 15.2.0 hc03a8fd_7
libuuid 1.41.5 h5eee18b_0
libxcb 1.17.0 h9b100fa_0
libzlib 1.3.1 h47b2149_1
ncurses 6.5 h7934f7d_0
openssl 3.5.6 h1b28b03_0
packaging 26.0 py311h06a4308_0
pip 26.0.1 pyhc872135_1
pthread-stubs 0.3 h0ce48e5_1
python 3.11.15 h741d88c_0
readline 8.3 hc2a1206_0
setuptools 82.0.1 py311h06a4308_0
sqlite 3.51.2 h3e8d24a_0
tk 8.6.15 h54e0aa7_0
tzdata 2026a he532380_0
wheel 0.46.3 py311h06a4308_0
xorg-libx11 1.8.12 h9b100fa_1
xorg-libxau 1.0.12 h9b100fa_0
xorg-libxdmcp 1.1.5 h9b100fa_0
xorg-xorgproto 2024.1 h5eee18b_1
xz 5.8.2 h448239c_0
zlib 1.3.1 h47b2149_1
(omm-pip) jag@jag:~$

[peastman]

Sorry, that should be openmmml with no space. See the installation instructions.

[jagriti-iisc]

pip install openmm[cuda12] openmmml mace-torch installs torch 2.11.0+cu130 (CUDA 13), but my driver (570.195.03) only supports up to CUDA 12.8. CUDA is not available. Should I force a CUDA 12 PyTorch with pip install torch --index-url https://download.pytorch.org/whl/cu128 before installing mace-torch?

"
PyTorch: 2.11.0+cu130
/home/jag/anaconda3/envs/omm-pip/lib/python3.11/site-packages/torch/cuda/init.py:180: UserWarning: CUDA initialization: The NVIDIA driver on your system is too old (found version 12080). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver. (Triggered internally at /pytorch/c10/cuda/CUDAFunctions.cpp:119.)
return torch._C._cuda_getDeviceCount() > 0
CUDA: False
Traceback (most recent call last):
File "", line 7, in
File "/home/jag/anaconda3/envs/omm-pip/lib/python3.11/site-packages/torch/cuda/init.py", line 653, in get_device_name
return get_device_properties(device).name
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/jag/anaconda3/envs/omm-pip/lib/python3.11/site-packages/torch/cuda/init.py", line 686, in get_device_properties
_lazy_init() # will define _get_device_properties
^^^^^^^^^^^^
File "/home/jag/anaconda3/envs/omm-pip/lib/python3.11/site-packages/torch/cuda/init.py", line 478, in _lazy_init
torch._C._cuda_init()
RuntimeError: The NVIDIA driver on your system is too old (found version 12080). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver.

[peastman]

Can you upgrade your driver? If so that will be the easiest.

[jagriti-iisc]

My driver is 570.195.03 which supports up to CUDA 12.8. I'd prefer not to update the driver as I have several working production environments that I don't want to risk breaking. Is there a way to force pip to install a CUDA 12 compatible PyTorch instead? For example: pip install torch --index-url https://download.pytorch.org/whl/cu128 before installing mace-torch?

[peastman]

You could install an older version of PyTorch that was compatible with it. I can't promise the other packages you need to install will be compatible with that old PyTorch though.

[jagriti-iisc]

i updated the cuda tookit, how can we install plumed into this env?
(openmm) jag@jag:~$ conda install -c conda-forge openmm-plumed py-plumed -y
Collecting package metadata (current_repodata.json): done
Solving environment: unsuccessful initial attempt using frozen solve. Retrying with flexible solve.
Solving environment: unsuccessful attempt using repodata from current_repodata.json, retrying with next repodata source.
Collecting package metadata (repodata.json): done
Solving environment: unsuccessful initial attempt using frozen solve. Retrying with flexible solve.
Solving environment: -
Found conflicts! Looking for incompatible packages.
This can take several minutes. Press CTRL-C to abort.
failed

UnsatisfiableError: The following specifications were found
to be incompatible with the existing python installation in your environment:

Specifications:

• openmm-plumed -> python[version='3.10.|3.11.|3.9.|3.8.|3.7.|3.6.']
• py-plumed -> python[version='3.10.|3.11.|3.9.|3.8.|3.7.|3.6.']
• py-plumed -> python[version='>=2.7,<2.8.0a0|>=3.10,<3.11.0a0|>=3.11,<3.12.0a0|>=3.9,<3.10.0a0|>=3.13.0rc1,<3.14.0a0|>=3.12,<3.13.0a0|>=3.8,<3.9.0a0|>=3.12.0rc3,<3.13.0a0|>=3.7,<3.8.0a0|>=3.6,<3.7.0a0']

Your python: python=3.14

If python is on the left-most side of the chain, that's the version you've asked for.
When python appears to the right, that indicates that the thing on the left is somehow
not available for the python version you are constrained to. Note that conda will not
change your python version to a different minor version unless you explicitly specify
that.

The following specifications were found to be incompatible with each other:

Output in format: Requested package -> Available versions

Package zstd conflicts for:
python=3.14 -> zstd[version='>=1.5.6,<1.6.0a0|>=1.5.7,<1.6.0a0']
openmm-plumed -> python[version='>=3.14,<3.15.0a0'] -> zstd[version='>=1.5.6,<1.6.0a0|>=1.5.7,<1.6.0a0']

Package pypy3.9 conflicts for:
py-plumed -> pypy3.9[version='>=7.3.11|>=7.3.12|>=7.3.15|>=7.3.9']
py-plumed -> python[version='>=3.9,<3.10.0a0'] -> pypy3.9[version='7.3.|7.3.11.|7.3.12.|7.3.13.|7.3.15.|7.3.9.|7.3.8.*']

Package xdrfile conflicts for:
openmm-plumed -> plumed[version='>=2.7,<2.8.0a0'] -> xdrfile[version='>=1.1.4,<1.1.5.0a0']
py-plumed -> plumed -> xdrfile[version='>=1.1.4,<1.1.5.0a0']

Package pypy3.8 conflicts for:
py-plumed -> python[version='>=3.8,<3.9.0a0'] -> pypy3.8[version='7.3.|7.3.11.|7.3.9.|7.3.8.']
py-plumed -> pypy3.8[version='>=7.3.11|>=7.3.9']

Package pypy3.7 conflicts for:
py-plumed -> pypy3.7[version='>=7.3.3|>=7.3.5|>=7.3.7']
py-plumed -> python[version='>=3.7,<3.8.0a0'] -> pypy3.7[version='7.3.|7.3.3.|7.3.4.|7.3.5.|7.3.7.*']

Package ca-certificates conflicts for:
python=3.14 -> openssl[version='>=3.5.6,<4.0a0'] -> ca-certificates
py-plumed -> python[version='>=2.7,<2.8.0a0'] -> ca-certificatesThe following specifications were found to be incompatible with your system:

• feature:/linux-64::__glibc==2.35=0
• feature:|@/linux-64::__glibc==2.35=0
• openmm-plumed -> __glibc[version='>=2.17|>=2.17,<3.0.a0|>=2.28,<3.0.a0']
• openmm-plumed -> sysroot_linux-64[version='>=2.17'] -> __glibc[version='>=2.28|>=2.34|>=2.39']
• py-plumed -> __glibc[version='>=2.17,<3.0.a0']
• py-plumed -> libgcc[version='>=13'] -> __glibc[version='>=2.17|>=2.28,<3.0.a0']

Your installed version is: 2.35

Note that strict channel priority may have removed packages required for satisfiability.

(openmm) jag@jag:~$

[peastman]

I don't advise it. Plumed is only available through conda, not pip. MACE is only available through pip, not conda. And mixing both package managers in the same environment is a great way to end up with a broken environment.

[jagriti-iisc]

so how to run a biased md simuation openmm ?

[peastman]

What kind of bias do you want? It supports lots of types of biasing forces.