Simulations don't run when `nvidia-smi` is unavailable

[david-zwicker]

I tried running simulations using the torch backend on my mac laptop, which does not have an Nvidia GPU. The simulation then fails because the profiler does not work (see error message below). It would be great if the package also worked with just the cpu device of torch. In any case, the hard-coded use of nvidia-smi implicitly assumes that torch uses the cuda backend, which is not general (and might led to problems in the future when other backends become more common).

[...]

File ~/user/miniconda3/envs/voxenv/lib/python3.12/site-packages/evoxels/solvers.py:122, in TimeDependentSolver._handle_outputs(self, u, frame, time, slice_idx, vtk_out, verbose, plot_bounds, colormap)
    119     self.vf.fields[name] = self.vg.export_scalar_field_to_numpy(u_out[i:i+1])
    121 if verbose:
--> [122](https://file+.vscode-resource.vscode-cdn.net/Users/dzwicker/Documents/Referee%20Reports/2026-01_JOSS/evoxels/docs/notebooks/~/user/miniconda3/envs/voxenv/lib/python3.12/site-packages/evoxels/solvers.py:122)     self.profiler.update_memory_stats()
    124 if self.vg.lib.isnan(u_out).any():
    125     print(f"NaN detected in frame {frame} at time {time}. Aborting simulation.")

File ~/user/miniconda3/envs/voxenv/lib/python3.12/site-packages/evoxels/profiler.py:27, in MemoryProfiler.update_memory_stats(self)
     25 self.max_used_cpu = np.max((self.max_used_cpu, used_cpu))
     26 used = self.get_cuda_memory_from_nvidia_smi()
---> [27](https://file+.vscode-resource.vscode-cdn.net/Users/dzwicker/Documents/Referee%20Reports/2026-01_JOSS/evoxels/docs/notebooks/~/user/miniconda3/envs/voxenv/lib/python3.12/site-packages/evoxels/profiler.py:27) self.max_used_gpu = np.max((self.max_used_gpu, used))

File ~/user/miniconda3/envs/voxenv/lib/python3.12/site-packages/numpy/_core/fromnumeric.py:3123, in max(a, axis, out, keepdims, initial, where)
   3011 @array_function_dispatch(_max_dispatcher)
   3012 @set_module('numpy')
   3013 def max(a, axis=None, out=None, keepdims=np._NoValue, initial=np._NoValue,
   3014          where=np._NoValue):
   3015     """
   3016     Return the maximum of an array or maximum along an axis.
   3017 
   (...)   3121     5
   3122     """
-> [3123](https://file+.vscode-resource.vscode-cdn.net/Users/dzwicker/Documents/Referee%20Reports/2026-01_JOSS/evoxels/docs/notebooks/~/user/miniconda3/envs/voxenv/lib/python3.12/site-packages/numpy/_core/fromnumeric.py:3123)     return _wrapreduction(a, np.maximum, 'max', axis, None, out,
   3124                           keepdims=keepdims, initial=initial, where=where)

File ~/user/miniconda3/envs/voxenv/lib/python3.12/site-packages/numpy/_core/fromnumeric.py:83, in _wrapreduction(obj, ufunc, method, axis, dtype, out, **kwargs)
     80         else:
     81             return reduction(axis=axis, out=out, **passkwargs)
---> [83](https://file+.vscode-resource.vscode-cdn.net/Users/dzwicker/Documents/Referee%20Reports/2026-01_JOSS/evoxels/docs/notebooks/~/user/miniconda3/envs/voxenv/lib/python3.12/site-packages/numpy/_core/fromnumeric.py:83) return ufunc.reduce(obj, axis, dtype, out, **passkwargs)

TypeError: '>=' not supported between instances of 'int' and 'NoneType'

[daubners]

@david-zwicker Good point. The main reason I am using the nvidi-smi command is that jax does not come with the same level of gpu ram information as torch. Also, there are some funky things like

import torch
free, total = torch.cuda.mem_get_info('cuda')
print((total-free) / 1024**2)

is actually not the same as the MB reported by nvidia-smi.

I've now wrapped it in a way that even if you're using a GPU which is not cuda, the code should run. (see #33 )