Installation¶

Install torch-fem with the notebook extra to get Jupyter widgets, interactive PyVista rendering, and animation support, which the example notebooks need:

pip install torch-fem[notebook]

Plain pip install torch-fem installs only what is required to build and solve models, if you do not intend to run the notebooks.

GPU support (optional)¶

Install PyTorch for your CUDA version. For CUDA 13:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

For CUDA 12, use the cu129 index URL instead. CUDA 11 is not supported.

Advanced solver backends (optional)¶

torch-fem runs its iterative solvers in PyTorch itself, on either device, and falls back to SciPy for a direct solve and for the algebraic multigrid preconditioner on CPU. Two optional backends can be faster, and torch-fem picks up either one as soon as it is installed.

Pardiso is a direct CPU solver:

pip install pypardiso

Once installed, torch-fem uses it automatically for a direct solve, which is what small models get by default.

AmgX is NVIDIA's GPU algebraic multigrid library, which needs far fewer iterations than a Jacobi preconditioner. It ships no wheels, so build it from source against a matching CUDA toolkit by following the instructions in the AmgX repository, then point torch-fem at the resulting shared library:

export AMGX_DLL=/path/to/libamgxsh.so # amgxsh.dll on Windows

Iterative solves on the GPU then use it automatically. Pass preconditioner="jacobi" to fall back to the diagonal one.

Development (optional)¶

To develop new features for torch-fem, you should fork the GitHub repository and clone it to your machine via

git clone <repo_url>

In the root of the cloned project, make the installation editable as

pip install -e ".[all]"

The all extra combines the notebook extra with the development tools (pytest, testbook, ruff) in the dev extra. Both are required to run the test suite, which executes the example notebooks.

Now, the package is linked to this local directory and whenever you use import torchfem, it will use the latest code.