transport.load_cpu
def load_cpu() -> NoneRegister ``torch.ops.transport.*`` (CPU backend).
Always registers the Python-side Sinkhorn custom_ops; additionally
loads the C++ ``exact_emd`` extension when a prebuilt ``.so`` is
shipped in the wheel or a C++ toolchain is available for the JIT
fallback. With ``TORCHMATCH_SKIP_TRANSPORT=1`` at build time and no
toolchain at install, the extension is absent and EXACT_EMD fails
with a targeted RuntimeError at call time; the Sinkhorn backends
keep working. Set ``TORCHMATCH_DEBUG_LOADER=1`` to surface the
underlying JIT failure when diagnosing a partially-broken extension.
Source: torchmatch/transport/matrix/_cpu.py:55