Dependency search

In most cases, there is no need to manually obtain the dependent libraries/packages. The build system will automatically search for the required packages at usual locations. If the package is already installed on the system, it will be used to build NEML2. Otherwise, a compatible version of the package will be downloaded and installed under the NEML2 build directory.

In case the package of interest has been installed at a non-conventional location, and CMake's default searching mechanism fails to find it, some special configure options can be used to help locate it. For a package named <PackageName>, the following variables are tried in sequence:

• <PackageName>_ROOT CMake variable
• <PACKAGENAME>_ROOT CMake variable
• <PackageName>_ROOT enviroment variable
• <PACKAGENAME>_ROOT environment variable

Please refer to the CMake documentation for additional hints that can be used to facilitate the package search procedure.

List of dependencies

C++ backend:

• PyTorch, version 2.8.0.
• WASP as the lexing and parsing backend for HIT.
• Testing:
  • Catch2 for unit and regression testing.

The Runner:

• argparse for command-line argument parsing.
• Profiling:
  • Gperftools for profiling purposes.

Python package:

• Python development libraries for python bindings.
• pybind11-stubgen for extracting stubs from Python bindings.
• pyzag for training material models.
• Testing:
  • pytest for testing Python bindings.

Documentation:

• Doxygen for building the documentation.
• Doxygen Awesome the documentation theme.
• graphviz for model visualization.
• PyYAML for extracting syntax documentation.

Work dispatcher:

• TIMPI for coordinating parallel workers.
• json for outputting event traces.

CSV parser:

• CSV Parser for reading CSV files

In addition to standard system library locations, the CMake configure script also searches for an installed torch Python package. Recent PyTorch releases within a few minor versions are likely to be compatible.

Warning

If no PyTorch is found after searching, a CPU-only libtorch binary is downloaded from the official website. Such libtorch is not able to use the CUDA co-processor, even if there is one. If using CUDA is desired, please install the torch Python package or a CUDA-enabled libtorch before configuring NEML2.

Note

We strive to keep up with the rapid development of PyTorch. The NEML2 PyTorch dependency is updated on a quarterly basis. If there is a particular version of PyTorch you'd like to use which is found to be incompatible with NEML2, please feel free to create an issue.

Skipping dependencies

In some cases, certain dependencies cannot be obtained or are incompatible with the build system, and it becomes desirable to keep using NEML2 with some capabilities disabled.

The following table summarizes the configure options that determine when a dependency is required, and hence how a dependency can be skipped.

Dependency	Dependent configure option(s)
Torch
WASP
Catch2	NEML2_TESTS
argparse	NEML2_RUNNER
Gperftools	NEML2_RUNNER && CMAKE_BUILD_TYPE == Profiling
Python development libraries	NEML2_PYBIND
pybind11-stubgen	NEML2_PYBIND
pyzag	NEML2_PYBIND && NEML2_TESTS
pytest	NEML2_PYBIND && NEML2_TESTS
Doxygen	NEML2_DOC
Doxygen Awesome	NEML2_DOC
graphviz	NEML2_PYBIND && NEML2_TESTS
PyYAML	NEML2_DOC
MPI, TIMPI	NEML2_WORK_DISPATCHER
json	NEML2_JSON
CSV Parser	NEML2_CSV

Dependency search

In most cases, there is no need to manually obtain the dependent libraries/packages. The build system will automatically search for the required packages at usual locations. If the package is already installed on the system, it will be used to build NEML2. Otherwise, a compatible version of the package will be downloaded and installed under the NEML2 build directory.

In case the package of interest has been installed at a non-conventional location, and CMake's default searching mechanism fails to find it, some special configure options can be used to help locate it. For a package named <PackageName>, the following variables are tried in sequence:

• <PackageName>_ROOT CMake variable
• <PACKAGENAME>_ROOT CMake variable
• <PackageName>_ROOT enviroment variable
• <PACKAGENAME>_ROOT environment variable

Please refer to the CMake documentation for additional hints that can be used to facilitate the package search procedure.

List of dependencies

C++ backend:

• PyTorch, version 2.8.0.
• WASP as the lexing and parsing backend for HIT.
• Testing:
  • Catch2 for unit and regression testing.

The Runner:

• argparse for command-line argument parsing.
• Profiling:
  • Gperftools for profiling purposes.

Python package:

• Python development libraries for python bindings.
• pybind11-stubgen for extracting stubs from Python bindings.
• pyzag for training material models.
• Testing:
  • pytest for testing Python bindings.

Documentation:

• Doxygen for building the documentation.
• Doxygen Awesome the documentation theme.
• graphviz for model visualization.
• PyYAML for extracting syntax documentation.

Work dispatcher:

• TIMPI for coordinating parallel workers.
• json for outputting event traces.

CSV parser:

• CSV Parser for reading CSV files

In addition to standard system library locations, the CMake configure script also searches for an installed torch Python package. Recent PyTorch releases within a few minor versions are likely to be compatible.

Warning

If no PyTorch is found after searching, a CPU-only libtorch binary is downloaded from the official website. Such libtorch is not able to use the CUDA co-processor, even if there is one. If using CUDA is desired, please install the torch Python package or a CUDA-enabled libtorch before configuring NEML2.

Note

We strive to keep up with the rapid development of PyTorch. The NEML2 PyTorch dependency is updated on a quarterly basis. If there is a particular version of PyTorch you'd like to use which is found to be incompatible with NEML2, please feel free to create an issue.

Skipping dependencies

In some cases, certain dependencies cannot be obtained or are incompatible with the build system, and it becomes desirable to keep using NEML2 with some capabilities disabled.

The following table summarizes the configure options that determine when a dependency is required, and hence how a dependency can be skipped.

Dependency	Dependent configure option(s)
Torch
WASP
Catch2	NEML2_TESTS
argparse	NEML2_RUNNER
Gperftools	NEML2_RUNNER && CMAKE_BUILD_TYPE == Profiling
Python development libraries	NEML2_PYBIND
pybind11-stubgen	NEML2_PYBIND
pyzag	NEML2_PYBIND && NEML2_TESTS
pytest	NEML2_PYBIND && NEML2_TESTS
Doxygen	NEML2_DOC
Doxygen Awesome	NEML2_DOC
graphviz	NEML2_PYBIND && NEML2_TESTS
PyYAML	NEML2_DOC
MPI, TIMPI	NEML2_WORK_DISPATCHER
json	NEML2_JSON
CSV Parser	NEML2_CSV