pyPFC logo

Troubleshooting¶

If you encounter installation issues or errors when trying to run pyPFC, answers can hopefully be found below:

I have a Nvidia GPU installed but pyPFC tells me I don't¶

Ensure that PyTorch is installed with CUDA enabled. To be sure, you can also check your setup by running:

import torch
print(torch.cuda.is_available())
print(torch.cuda.device_count())
print(torch.version.cuda)

If is_available() is False or device_count() is 0, PyTorch cannot see your GPU. See the section on GPU support in the Installation Guide on how to enable CUDA.

The solid crystal phase fails to stabilize and/or appears to "melt" away¶

This is most likely due to domain_size not being set correctly to accommodate the current lattice periodicity.

The atom positions obtained by `interpolate_density_maxima` do not seem to coincide with the density field maxima¶

This is likely related to either insufficient grid resolution or too low interpolation order. The former issue is mitigated by reducing the values in ddiv and in the latter case density_interp_order should be increased. Usually density_interp_order=2 is fine and increasing the number will also increase the time spent on interpolation.

ImportError: No module named 'pypfc'¶

Make sure you're in the correct environment and pyPFC is installed:

pip list | grep pypfc

CUDA Out of Memory¶

Reduce simulation size or use CPU:

params = {'device_type': 'CPU'}
domain_size = [size_x, size_y, size_z] # Consider reducing these numbers
pfc.setup_simulation(domain_size, config=params)

FFT Errors¶

Ensure grid dimensions are even numbers (pyPFC should issue a warning if this is not the case):

# Good: even dimensions
ndiv = [64, 64, 32]

# Bad: odd dimensions
ndiv = [63, 65, 31]

Performance Issues¶

Slow Simulations¶

Enable GPU: Verify CUDA installation and GPU detection
Grid Size: Start with smaller grids for testing

Memory Issues¶

Reduce Grid Size: Use a smaller simulation domain or reduce the number of grid points ndiv
Use Single Precision: Set dtype_gpu: 'torch.float32' and dtype_cpu': 'np.single'
Clear Device Cache: Call torch.cuda.empty_cache()