torchmatch
function

assignment.load_cuda

def load_cuda() -> None

Register the CUDA-only ops in ``torch.ops.assignment``.

Adds Munkres classical (``munkres``), the experimental Munkres hybrid (``hybrid``), Lawler tree-augmentation (``lawler``), and the CUDA backend of ``jonker_dense_batch``. Prefers a prebuilt extension shipped in the wheel and falls back to JIT-compiling the C++/CUDA sources via :func:`torch.utils.cpp_extension.load`. Set ``TORCHMATCH_FORCE_JIT=1`` to skip the prebuilt path.

Source: torchmatch/assignment/_cuda.py:61